/Users/deen/code/yugabyte-db/src/yb/yql/pgwrapper/pg_index_backfill-test.cc
Line | Count | Source (jump to first uncovered line) |
1 | | // Copyright (c) YugaByte, Inc. |
2 | | // |
3 | | // Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except |
4 | | // in compliance with the License. You may obtain a copy of the License at |
5 | | // |
6 | | // http://www.apache.org/licenses/LICENSE-2.0 |
7 | | // |
8 | | // Unless required by applicable law or agreed to in writing, software distributed under the License |
9 | | // is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express |
10 | | // or implied. See the License for the specific language governing permissions and limitations |
11 | | // under the License. |
12 | | |
13 | | #include <cmath> |
14 | | #include <map> |
15 | | #include <string> |
16 | | #include <vector> |
17 | | |
18 | | #include "yb/client/table_info.h" |
19 | | |
20 | | #include "yb/common/schema.h" |
21 | | |
22 | | #include "yb/integration-tests/backfill-test-util.h" |
23 | | |
24 | | #include "yb/util/async_util.h" |
25 | | #include "yb/util/backoff_waiter.h" |
26 | | #include "yb/util/format.h" |
27 | | #include "yb/util/monotime.h" |
28 | | #include "yb/util/status_format.h" |
29 | | #include "yb/util/test_thread_holder.h" |
30 | | #include "yb/util/test_util.h" |
31 | | #include "yb/util/tsan_util.h" |
32 | | |
33 | | #include "yb/yql/pgwrapper/libpq_test_base.h" |
34 | | #include "yb/yql/pgwrapper/libpq_utils.h" |
35 | | |
36 | | using namespace std::chrono_literals; |
37 | | |
38 | | namespace yb { |
39 | | namespace pgwrapper { |
40 | | |
41 | | namespace { |
42 | | |
43 | | constexpr auto kColoDbName = "colodb"; |
44 | | constexpr auto kDatabaseName = "yugabyte"; |
45 | | constexpr auto kIndexName = "iii"; |
46 | | constexpr auto kTableName = "ttt"; |
47 | | const client::YBTableName kYBTableName(YQLDatabase::YQL_DATABASE_PGSQL, kDatabaseName, kTableName); |
48 | | |
49 | | } // namespace |
50 | | |
51 | | YB_DEFINE_ENUM(IndexStateFlag, (kIndIsLive)(kIndIsReady)(kIndIsValid)); |
52 | | typedef EnumBitSet<IndexStateFlag> IndexStateFlags; |
53 | | |
54 | | class PgIndexBackfillTest : public LibPqTestBase { |
55 | | public: |
56 | 0 | void SetUp() override { |
57 | 0 | LibPqTestBase::SetUp(); |
58 | |
|
59 | 0 | conn_ = std::make_unique<PGConn>(ASSERT_RESULT(ConnectToDB(kDatabaseName))); |
60 | 0 | } |
61 | | |
62 | 0 | void UpdateMiniClusterOptions(ExternalMiniClusterOptions* options) override { |
63 | 0 | options->extra_master_flags.push_back("--ysql_disable_index_backfill=false"); |
64 | 0 | options->extra_master_flags.push_back( |
65 | 0 | Format("--ysql_num_shards_per_tserver=$0", kTabletsPerServer)); |
66 | 0 | options->extra_tserver_flags.push_back("--ysql_disable_index_backfill=false"); |
67 | 0 | options->extra_tserver_flags.push_back( |
68 | 0 | Format("--ysql_num_shards_per_tserver=$0", kTabletsPerServer)); |
69 | 0 | } |
70 | | |
71 | | protected: |
72 | | bool HasClientTimedOut(const Status& s); |
73 | | void TestSimpleBackfill(const std::string& table_create_suffix = ""); |
74 | | void TestLargeBackfill(const int num_rows); |
75 | | void TestRetainDeleteMarkers(const std::string& db_name); |
76 | | const int kTabletsPerServer = 8; |
77 | | |
78 | | std::unique_ptr<PGConn> conn_; |
79 | | }; |
80 | | |
81 | | namespace { |
82 | | |
83 | | // A copy of the same function in pg_libpq-test.cc. Eventually, issue #6868 should provide a way to |
84 | | // do this easily for both this file and that. |
85 | | Result<string> GetTableIdByTableName( |
86 | 0 | client::YBClient* client, const string& namespace_name, const string& table_name) { |
87 | 0 | const auto tables = VERIFY_RESULT(client->ListTables()); |
88 | 0 | for (const auto& t : tables) { |
89 | 0 | if (t.namespace_name() == namespace_name && t.table_name() == table_name) { |
90 | 0 | return t.table_id(); |
91 | 0 | } |
92 | 0 | } |
93 | 0 | return STATUS(NotFound, "The table does not exist"); |
94 | 0 | } |
95 | | |
96 | 0 | Result<int> TotalBackfillRpcMetric(ExternalMiniCluster* cluster, const char* type) { |
97 | 0 | int total_rpc_calls = 0; |
98 | 0 | constexpr auto metric_name = "handler_latency_yb_tserver_TabletServerAdminService_BackfillIndex"; |
99 | 0 | for (auto ts : cluster->tserver_daemons()) { |
100 | 0 | auto val = VERIFY_RESULT(ts->GetInt64Metric("server", "yb.tabletserver", metric_name, type)); |
101 | 0 | total_rpc_calls += val; |
102 | 0 | VLOG(1) << ts->bind_host() << " for " << type << " returned " << val; |
103 | 0 | } |
104 | 0 | return total_rpc_calls; |
105 | 0 | } |
106 | | |
107 | 0 | Result<int> TotalBackfillRpcCalls(ExternalMiniCluster* cluster) { |
108 | 0 | return TotalBackfillRpcMetric(cluster, "total_count"); |
109 | 0 | } |
110 | | |
111 | 0 | Result<double> AvgBackfillRpcLatencyInMicros(ExternalMiniCluster* cluster) { |
112 | 0 | auto num_calls = VERIFY_RESULT(TotalBackfillRpcMetric(cluster, "total_count")); |
113 | 0 | double total_latency = VERIFY_RESULT(TotalBackfillRpcMetric(cluster, "total_sum")); |
114 | 0 | return total_latency / num_calls; |
115 | 0 | } |
116 | | |
117 | | } // namespace |
118 | | |
119 | 0 | bool PgIndexBackfillTest::HasClientTimedOut(const Status& s) { |
120 | 0 | if (!s.IsNetworkError()) { |
121 | 0 | return false; |
122 | 0 | } |
123 | | |
124 | | // The client timeout is set using the same backfill_index_client_rpc_timeout_ms for |
125 | | // postgres-tserver RPC and tserver-master RPC. Since they are the same value, it _may_ be |
126 | | // possible for either timeout message to show up, so accept either, even though the |
127 | | // postgres-tserver timeout is far more likely to show up. |
128 | | // |
129 | | // The first is postgres-tserver; the second is tserver-master. |
130 | 0 | const std::string msg = s.message().ToBuffer(); |
131 | 0 | return msg.find("Timed out: BackfillIndex RPC") != std::string::npos || |
132 | 0 | msg.find("Timed out waiting for Backfill Index") != std::string::npos; |
133 | 0 | } |
134 | | |
135 | 0 | void PgIndexBackfillTest::TestSimpleBackfill(const std::string& table_create_suffix) { |
136 | 0 | ASSERT_OK(conn_->ExecuteFormat( |
137 | 0 | "CREATE TABLE $0 (c char, i int, p point) $1", |
138 | 0 | kTableName, |
139 | 0 | table_create_suffix)); |
140 | 0 | ASSERT_OK(conn_->ExecuteFormat("INSERT INTO $0 VALUES ('a', 0, '(1, 2)')", kTableName)); |
141 | 0 | ASSERT_OK(conn_->ExecuteFormat("INSERT INTO $0 VALUES ('y', -5, '(0, -2)')", kTableName)); |
142 | 0 | ASSERT_OK(conn_->ExecuteFormat("INSERT INTO $0 VALUES ('b', 100, '(868, 9843)')", kTableName)); |
143 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE INDEX ON $0 (c ASC)", kTableName)); |
144 | | |
145 | | // Index scan to verify contents of index table. |
146 | 0 | const std::string query = Format("SELECT * FROM $0 ORDER BY c", kTableName); |
147 | 0 | ASSERT_TRUE(ASSERT_RESULT(conn_->HasIndexScan(query))); |
148 | 0 | PGResultPtr res = ASSERT_RESULT(conn_->Fetch(query)); |
149 | 0 | ASSERT_EQ(PQntuples(res.get()), 3); |
150 | 0 | ASSERT_EQ(PQnfields(res.get()), 3); |
151 | 0 | std::array<int, 3> values = { |
152 | 0 | ASSERT_RESULT(GetInt32(res.get(), 0, 1)), |
153 | 0 | ASSERT_RESULT(GetInt32(res.get(), 1, 1)), |
154 | 0 | ASSERT_RESULT(GetInt32(res.get(), 2, 1)), |
155 | 0 | }; |
156 | 0 | ASSERT_EQ(values[0], 0); |
157 | 0 | ASSERT_EQ(values[1], 100); |
158 | 0 | ASSERT_EQ(values[2], -5); |
159 | 0 | } |
160 | | |
161 | | // Checks that retain_delete_markers is false after index creation. |
162 | 0 | void PgIndexBackfillTest::TestRetainDeleteMarkers(const std::string& db_name) { |
163 | 0 | auto client = ASSERT_RESULT(cluster_->CreateClient()); |
164 | |
|
165 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE TABLE $0 (i int)", kTableName)); |
166 | 0 | const auto index_name = "ttt_idx"; |
167 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE INDEX $0 ON $1 (i ASC)", index_name, kTableName)); |
168 | | |
169 | | // Verify that retain_delete_markers was set properly in the index table schema. |
170 | 0 | const std::string table_id = ASSERT_RESULT(GetTableIdByTableName( |
171 | 0 | client.get(), db_name, index_name)); |
172 | 0 | auto table_info = std::make_shared<client::YBTableInfo>(); |
173 | 0 | { |
174 | 0 | Synchronizer sync; |
175 | 0 | ASSERT_OK(client->GetTableSchemaById(table_id, table_info, sync.AsStatusCallback())); |
176 | 0 | ASSERT_OK(sync.Wait()); |
177 | 0 | } |
178 | |
|
179 | 0 | ASSERT_EQ(table_info->schema.version(), 0); |
180 | 0 | ASSERT_FALSE(table_info->schema.table_properties().retain_delete_markers()); |
181 | 0 | } |
182 | | |
183 | 0 | void PgIndexBackfillTest::TestLargeBackfill(const int num_rows) { |
184 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE TABLE $0 (i int)", kTableName)); |
185 | | |
186 | | // Insert bunch of rows. |
187 | 0 | ASSERT_OK(conn_->ExecuteFormat( |
188 | 0 | "INSERT INTO $0 VALUES (generate_series(1, $1))", |
189 | 0 | kTableName, |
190 | 0 | num_rows)); |
191 | | |
192 | | // Create index. |
193 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE INDEX ON $0 (i ASC)", kTableName)); |
194 | | |
195 | | // All rows should be in the index. |
196 | 0 | const std::string query = Format( |
197 | 0 | "SELECT COUNT(*) FROM $0 WHERE i > 0", |
198 | 0 | kTableName); |
199 | 0 | ASSERT_TRUE(ASSERT_RESULT(conn_->HasIndexScan(query))); |
200 | 0 | PGResultPtr res = ASSERT_RESULT(conn_->Fetch(query)); |
201 | 0 | ASSERT_EQ(PQntuples(res.get()), 1); |
202 | 0 | ASSERT_EQ(PQnfields(res.get()), 1); |
203 | 0 | auto actual_num_rows = ASSERT_RESULT(GetInt64(res.get(), 0, 0)); |
204 | 0 | ASSERT_EQ(actual_num_rows, num_rows); |
205 | 0 | } |
206 | | |
207 | | // Make sure that backfill works. |
208 | 0 | TEST_F(PgIndexBackfillTest, YB_DISABLE_TEST_IN_TSAN(Simple)) { |
209 | 0 | TestSimpleBackfill(); |
210 | 0 | } |
211 | | |
212 | | // Make sure that partial indexes work for index backfill. |
213 | 0 | TEST_F(PgIndexBackfillTest, YB_DISABLE_TEST_IN_TSAN(Partial)) { |
214 | 0 | constexpr int kNumRows = 7; |
215 | |
|
216 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE TABLE $0 (i int, j int)", kTableName)); |
217 | 0 | ASSERT_OK(conn_->ExecuteFormat( |
218 | 0 | "INSERT INTO $0 VALUES (generate_series(1, $1), generate_series(-1, -$1, -1))", |
219 | 0 | kTableName, |
220 | 0 | kNumRows)); |
221 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE INDEX ON $0 (i ASC) WHERE j > -5", kTableName)); |
222 | | |
223 | | // Index scan to verify contents of index table. |
224 | 0 | { |
225 | 0 | const std::string query = Format("SELECT j FROM $0 WHERE j > -3 ORDER BY i", kTableName); |
226 | 0 | ASSERT_TRUE(ASSERT_RESULT(conn_->HasIndexScan(query))); |
227 | 0 | PGResultPtr res = ASSERT_RESULT(conn_->Fetch(query)); |
228 | 0 | ASSERT_EQ(PQntuples(res.get()), 2); |
229 | 0 | ASSERT_EQ(PQnfields(res.get()), 1); |
230 | 0 | std::array<int, 2> values = { |
231 | 0 | ASSERT_RESULT(GetInt32(res.get(), 0, 0)), |
232 | 0 | ASSERT_RESULT(GetInt32(res.get(), 1, 0)), |
233 | 0 | }; |
234 | 0 | ASSERT_EQ(values[0], -1); |
235 | 0 | ASSERT_EQ(values[1], -2); |
236 | 0 | } |
237 | 0 | { |
238 | 0 | const std::string query = Format( |
239 | 0 | "SELECT i FROM $0 WHERE j > -5 ORDER BY i DESC LIMIT 2", |
240 | 0 | kTableName); |
241 | 0 | ASSERT_TRUE(ASSERT_RESULT(conn_->HasIndexScan(query))); |
242 | 0 | PGResultPtr res = ASSERT_RESULT(conn_->Fetch(query)); |
243 | 0 | ASSERT_EQ(PQntuples(res.get()), 2); |
244 | 0 | ASSERT_EQ(PQnfields(res.get()), 1); |
245 | 0 | std::array<int, 2> values = { |
246 | 0 | ASSERT_RESULT(GetInt32(res.get(), 0, 0)), |
247 | 0 | ASSERT_RESULT(GetInt32(res.get(), 1, 0)), |
248 | 0 | }; |
249 | 0 | ASSERT_EQ(values[0], 4); |
250 | 0 | ASSERT_EQ(values[1], 3); |
251 | 0 | } |
252 | 0 | } |
253 | | |
254 | | // Make sure that expression indexes work for index backfill. |
255 | 0 | TEST_F(PgIndexBackfillTest, YB_DISABLE_TEST_IN_TSAN(Expression)) { |
256 | 0 | constexpr int kNumRows = 9; |
257 | |
|
258 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE TABLE $0 (i int, j int)", kTableName)); |
259 | 0 | ASSERT_OK(conn_->ExecuteFormat( |
260 | 0 | "INSERT INTO $0 VALUES (generate_series(1, $1), generate_series(11, 10 + $1))", |
261 | 0 | kTableName, |
262 | 0 | kNumRows)); |
263 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE INDEX ON $0 ((j % i))", kTableName)); |
264 | | |
265 | | // Index scan to verify contents of index table. |
266 | 0 | const std::string query = Format( |
267 | 0 | "SELECT j, i, j % i as mod FROM $0 WHERE j % i = 2 ORDER BY i", |
268 | 0 | kTableName); |
269 | 0 | ASSERT_TRUE(ASSERT_RESULT(conn_->HasIndexScan(query))); |
270 | 0 | PGResultPtr res = ASSERT_RESULT(conn_->Fetch(query)); |
271 | 0 | ASSERT_EQ(PQntuples(res.get()), 2); |
272 | 0 | ASSERT_EQ(PQnfields(res.get()), 3); |
273 | 0 | std::array<std::array<int, 3>, 2> values = {{ |
274 | 0 | { |
275 | 0 | ASSERT_RESULT(GetInt32(res.get(), 0, 0)), |
276 | 0 | ASSERT_RESULT(GetInt32(res.get(), 0, 1)), |
277 | 0 | ASSERT_RESULT(GetInt32(res.get(), 0, 2)), |
278 | 0 | }, |
279 | 0 | { |
280 | 0 | ASSERT_RESULT(GetInt32(res.get(), 1, 0)), |
281 | 0 | ASSERT_RESULT(GetInt32(res.get(), 1, 1)), |
282 | 0 | ASSERT_RESULT(GetInt32(res.get(), 1, 2)), |
283 | 0 | }, |
284 | 0 | }}; |
285 | 0 | ASSERT_EQ(values[0][0], 14); |
286 | 0 | ASSERT_EQ(values[0][1], 4); |
287 | 0 | ASSERT_EQ(values[0][2], 2); |
288 | 0 | ASSERT_EQ(values[1][0], 18); |
289 | 0 | ASSERT_EQ(values[1][1], 8); |
290 | 0 | ASSERT_EQ(values[1][2], 2); |
291 | 0 | } |
292 | | |
293 | | // Make sure that unique indexes work when index backfill is enabled. |
294 | 0 | TEST_F(PgIndexBackfillTest, YB_DISABLE_TEST_IN_TSAN(Unique)) { |
295 | 0 | constexpr int kNumRows = 3; |
296 | |
|
297 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE TABLE $0 (i int, j int)", kTableName)); |
298 | 0 | ASSERT_OK(conn_->ExecuteFormat( |
299 | 0 | "INSERT INTO $0 VALUES (generate_series(1, $1), generate_series(11, 10 + $1))", |
300 | 0 | kTableName, |
301 | 0 | kNumRows)); |
302 | | // Add row that would make j not unique. |
303 | 0 | ASSERT_OK(conn_->ExecuteFormat( |
304 | 0 | "INSERT INTO $0 VALUES (99, 11)", |
305 | 0 | kTableName, |
306 | 0 | kNumRows)); |
307 | | |
308 | | // Create unique index without failure. |
309 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE UNIQUE INDEX ON $0 (i ASC)", kTableName)); |
310 | | // Index scan to verify contents of index table. |
311 | 0 | const std::string query = Format( |
312 | 0 | "SELECT * FROM $0 ORDER BY i", |
313 | 0 | kTableName); |
314 | 0 | ASSERT_TRUE(ASSERT_RESULT(conn_->HasIndexScan(query))); |
315 | 0 | PGResultPtr res = ASSERT_RESULT(conn_->Fetch(query)); |
316 | 0 | ASSERT_EQ(PQntuples(res.get()), 4); |
317 | 0 | ASSERT_EQ(PQnfields(res.get()), 2); |
318 | | |
319 | | // Create unique index with failure. |
320 | 0 | Status status = conn_->ExecuteFormat("CREATE UNIQUE INDEX ON $0 (j ASC)", kTableName); |
321 | 0 | ASSERT_NOK(status); |
322 | 0 | const std::string msg = status.message().ToBuffer(); |
323 | 0 | ASSERT_TRUE(msg.find("duplicate key value violates unique constraint") != std::string::npos) |
324 | 0 | << status; |
325 | 0 | } |
326 | | |
327 | | // Make sure that indexes created in postgres nested DDL work and skip backfill (optimization). |
328 | 0 | TEST_F(PgIndexBackfillTest, YB_DISABLE_TEST_IN_TSAN(NestedDdl)) { |
329 | 0 | auto client = ASSERT_RESULT(cluster_->CreateClient()); |
330 | 0 | constexpr int kNumRows = 3; |
331 | |
|
332 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE TABLE $0 (i int, j int, UNIQUE (j))", kTableName)); |
333 | | |
334 | | // Make sure that the index create was not multi-stage. |
335 | 0 | const std::string table_id = ASSERT_RESULT(GetTableIdByTableName( |
336 | 0 | client.get(), kDatabaseName, kTableName)); |
337 | 0 | std::shared_ptr<client::YBTableInfo> table_info = std::make_shared<client::YBTableInfo>(); |
338 | 0 | Synchronizer sync; |
339 | 0 | ASSERT_OK(client->GetTableSchemaById(table_id, table_info, sync.AsStatusCallback())); |
340 | 0 | ASSERT_OK(sync.Wait()); |
341 | 0 | ASSERT_EQ(table_info->schema.version(), 1); |
342 | |
|
343 | 0 | ASSERT_OK(conn_->ExecuteFormat( |
344 | 0 | "INSERT INTO $0 VALUES (generate_series(1, $1), generate_series(11, 10 + $1))", |
345 | 0 | kTableName, |
346 | 0 | kNumRows)); |
347 | | |
348 | | // Add row that violates unique constraint on j. |
349 | 0 | Status status = conn_->ExecuteFormat( |
350 | 0 | "INSERT INTO $0 VALUES (99, 11)", |
351 | 0 | kTableName, |
352 | 0 | kNumRows); |
353 | 0 | ASSERT_NOK(status); |
354 | 0 | const std::string msg = status.message().ToBuffer(); |
355 | 0 | ASSERT_TRUE(msg.find("duplicate key value") != std::string::npos) << status; |
356 | 0 | } |
357 | | |
358 | | // Make sure that drop index works when index backfill is enabled (skips online schema migration for |
359 | | // now) |
360 | 0 | TEST_F(PgIndexBackfillTest, YB_DISABLE_TEST_IN_TSAN(Drop)) { |
361 | 0 | constexpr int kNumRows = 5; |
362 | |
|
363 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE TABLE $0 (i int, j int)", kTableName)); |
364 | 0 | ASSERT_OK(conn_->ExecuteFormat( |
365 | 0 | "INSERT INTO $0 VALUES (generate_series(1, $1), generate_series(11, 10 + $1))", |
366 | 0 | kTableName, |
367 | 0 | kNumRows)); |
368 | | |
369 | | // Create index. |
370 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE INDEX $0 ON $1 (i ASC)", kIndexName, kTableName)); |
371 | | |
372 | | // Drop index. |
373 | 0 | ASSERT_OK(conn_->ExecuteFormat("DROP INDEX $0", kIndexName)); |
374 | | |
375 | | // Ensure index is not used for scan. |
376 | 0 | const std::string query = Format( |
377 | 0 | "SELECT * FROM $0 ORDER BY i", |
378 | 0 | kTableName); |
379 | 0 | ASSERT_FALSE(ASSERT_RESULT(conn_->HasIndexScan(query))); |
380 | 0 | } |
381 | | |
382 | | // Make sure deletes to nonexistent rows look like noops to clients. This may seem too obvious to |
383 | | // necessitate a test, but logic for backfill is special in that it wants nonexistent index deletes |
384 | | // to be applied for the backfill process to use them. This test guards against that logic being |
385 | | // implemented incorrectly. |
386 | 0 | TEST_F(PgIndexBackfillTest, YB_DISABLE_TEST_IN_TSAN(NonexistentDelete)) { |
387 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE TABLE $0 (i int PRIMARY KEY)", kTableName)); |
388 | | |
389 | | // Delete to nonexistent row should return no rows. |
390 | 0 | PGResultPtr res = ASSERT_RESULT(conn_->FetchFormat( |
391 | 0 | "DELETE FROM $0 WHERE i = 1 RETURNING i", |
392 | 0 | kTableName)); |
393 | 0 | ASSERT_EQ(PQntuples(res.get()), 0); |
394 | 0 | ASSERT_EQ(PQnfields(res.get()), 1); |
395 | 0 | } |
396 | | |
397 | | // Make sure that index backfill on large tables backfills all data. |
398 | 0 | TEST_F(PgIndexBackfillTest, YB_DISABLE_TEST_IN_TSAN(Large)) { |
399 | 0 | constexpr int kNumRows = 10000; |
400 | 0 | TestLargeBackfill(kNumRows); |
401 | 0 | auto expected_calls = cluster_->num_tablet_servers() * kTabletsPerServer; |
402 | 0 | auto actual_calls = ASSERT_RESULT(TotalBackfillRpcCalls(cluster_.get())); |
403 | 0 | ASSERT_GE(actual_calls, expected_calls); |
404 | 0 | } |
405 | | |
406 | | class PgIndexBackfillTestChunking : public PgIndexBackfillTest { |
407 | | protected: |
408 | 0 | void UpdateMiniClusterOptions(ExternalMiniClusterOptions* options) override { |
409 | 0 | PgIndexBackfillTest::UpdateMiniClusterOptions(options); |
410 | 0 | options->extra_tserver_flags.push_back( |
411 | 0 | Format("--TEST_backfill_paging_size=$0", kBatchSize)); |
412 | 0 | options->extra_tserver_flags.push_back( |
413 | 0 | Format("--backfill_index_write_batch_size=$0", kBatchSize)); |
414 | 0 | options->extra_tserver_flags.push_back( |
415 | 0 | Format("--ysql_prefetch_limit=$0", kPrefetchSize)); |
416 | 0 | } |
417 | | const int kBatchSize = 200; |
418 | | const int kPrefetchSize = 128; |
419 | | }; |
420 | | |
421 | | // Set batch size and prefetch limit such that: |
422 | | // Each tablet requires multiple RPC calls from the master to complete backfill. |
423 | | // Also, set the ysql_prefetch_size small to ensure that each of these |
424 | | // `BACKFILL INDEX` calls will fetch data from the tserver at least 2 times. |
425 | | // Fetch metrics to ensure that there have been > num_tablets rpc's. |
426 | | TEST_F_EX( |
427 | 0 | PgIndexBackfillTest, YB_DISABLE_TEST_IN_TSAN(BackfillInChunks), PgIndexBackfillTestChunking) { |
428 | 0 | constexpr int kNumRows = 10000; |
429 | 0 | TestLargeBackfill(kNumRows); |
430 | |
|
431 | 0 | const size_t effective_batch_size = |
432 | 0 | static_cast<size_t>(kPrefetchSize * ceil(1.0 * kBatchSize / kPrefetchSize)); |
433 | 0 | const size_t min_expected_calls = |
434 | 0 | static_cast<size_t>(ceil(1.0 * kNumRows / effective_batch_size)); |
435 | 0 | auto actual_calls = ASSERT_RESULT(TotalBackfillRpcCalls(cluster_.get())); |
436 | 0 | LOG(INFO) << "Had " << actual_calls << " backfill rpc calls. " |
437 | 0 | << "Expected at least " << kNumRows << "/" << effective_batch_size << " = " |
438 | 0 | << min_expected_calls; |
439 | 0 | ASSERT_GE(actual_calls, min_expected_calls); |
440 | 0 | } |
441 | | |
442 | | class PgIndexBackfillTestThrottled : public PgIndexBackfillTest { |
443 | | protected: |
444 | 0 | void UpdateMiniClusterOptions(ExternalMiniClusterOptions* options) override { |
445 | 0 | PgIndexBackfillTest::UpdateMiniClusterOptions(options); |
446 | 0 | options->extra_master_flags.push_back( |
447 | 0 | Format("--ysql_index_backfill_rpc_timeout_ms=$0", kBackfillRpcDeadlineLargeMs)); |
448 | |
|
449 | 0 | options->extra_tserver_flags.push_back("--ysql_prefetch_limit=100"); |
450 | 0 | options->extra_tserver_flags.push_back("--backfill_index_write_batch_size=100"); |
451 | 0 | options->extra_tserver_flags.push_back( |
452 | 0 | Format("--backfill_index_rate_rows_per_sec=$0", kBackfillRateRowsPerSec)); |
453 | 0 | options->extra_tserver_flags.push_back( |
454 | 0 | Format("--num_concurrent_backfills_allowed=$0", kNumConcurrentBackfills)); |
455 | 0 | } |
456 | | |
457 | | protected: |
458 | | const int kBackfillRateRowsPerSec = 100; |
459 | | const int kNumConcurrentBackfills = 1; |
460 | | const int kBackfillRpcDeadlineLargeMs = 10 * 60 * 1000; |
461 | | }; |
462 | | |
463 | | // Set the backfill batch size and backfill rate |
464 | | // Check that the time taken to backfill is no less than what is expected. |
465 | | TEST_F_EX( |
466 | 0 | PgIndexBackfillTest, YB_DISABLE_TEST_IN_TSAN(ThrottledBackfill), PgIndexBackfillTestThrottled) { |
467 | 0 | constexpr int kNumRows = 10000; |
468 | 0 | auto start_time = CoarseMonoClock::Now(); |
469 | 0 | TestLargeBackfill(kNumRows); |
470 | 0 | auto end_time = CoarseMonoClock::Now(); |
471 | 0 | auto expected_time = MonoDelta::FromSeconds( |
472 | 0 | kNumRows * 1.0 / |
473 | 0 | (cluster_->num_tablet_servers() * kNumConcurrentBackfills * kBackfillRateRowsPerSec)); |
474 | 0 | ASSERT_GE(MonoDelta{end_time - start_time}, expected_time); |
475 | | |
476 | | // Expect only 1 call per tablet |
477 | 0 | const size_t expected_calls = cluster_->num_tablet_servers() * kTabletsPerServer; |
478 | 0 | auto actual_calls = ASSERT_RESULT(TotalBackfillRpcCalls(cluster_.get())); |
479 | 0 | ASSERT_EQ(actual_calls, expected_calls); |
480 | |
|
481 | 0 | auto avg_rpc_latency_usec = ASSERT_RESULT(AvgBackfillRpcLatencyInMicros(cluster_.get())); |
482 | 0 | LOG(INFO) << "Avg backfill latency was " << avg_rpc_latency_usec << " us"; |
483 | 0 | ASSERT_LE(avg_rpc_latency_usec, kBackfillRpcDeadlineLargeMs * 1000); |
484 | 0 | } |
485 | | |
486 | | class PgIndexBackfillTestDeadlines : public PgIndexBackfillTest { |
487 | | protected: |
488 | 0 | void UpdateMiniClusterOptions(ExternalMiniClusterOptions* options) override { |
489 | 0 | options->extra_master_flags.push_back("--ysql_disable_index_backfill=false"); |
490 | 0 | options->extra_master_flags.push_back( |
491 | 0 | Format("--ysql_num_shards_per_tserver=$0", kTabletsPerServer)); |
492 | 0 | options->extra_master_flags.push_back( |
493 | 0 | Format("--ysql_index_backfill_rpc_timeout_ms=$0", kBackfillRpcDeadlineSmallMs)); |
494 | 0 | options->extra_master_flags.push_back( |
495 | 0 | Format("--backfill_index_timeout_grace_margin_ms=$0", kBackfillRpcDeadlineSmallMs / 2)); |
496 | |
|
497 | 0 | options->extra_tserver_flags.push_back("--ysql_disable_index_backfill=false"); |
498 | 0 | options->extra_tserver_flags.push_back( |
499 | 0 | Format("--ysql_num_shards_per_tserver=$0", kTabletsPerServer)); |
500 | 0 | options->extra_tserver_flags.push_back("--ysql_prefetch_limit=100"); |
501 | 0 | options->extra_tserver_flags.push_back("--backfill_index_write_batch_size=100"); |
502 | 0 | options->extra_tserver_flags.push_back( |
503 | 0 | Format("--backfill_index_rate_rows_per_sec=$0", kBackfillRateRowsPerSec)); |
504 | 0 | options->extra_tserver_flags.push_back( |
505 | 0 | Format("--num_concurrent_backfills_allowed=$0", kNumConcurrentBackfills)); |
506 | 0 | } |
507 | | |
508 | | protected: |
509 | | const int kBackfillRpcDeadlineSmallMs = 10000; |
510 | | const int kBackfillRateRowsPerSec = 100; |
511 | | const int kNumConcurrentBackfills = 1; |
512 | | const int kTabletsPerServer = 1; |
513 | | }; |
514 | | |
515 | | // Set the backfill batch size, backfill rate and a low timeout for backfill rpc. |
516 | | // Ensure that the backfill is completed. And that the avg rpc latency is |
517 | | // below what is set as the timeout. |
518 | | TEST_F_EX( |
519 | | PgIndexBackfillTest, |
520 | | YB_DISABLE_TEST_IN_TSAN(BackfillRespectsDeadline), |
521 | 0 | PgIndexBackfillTestDeadlines) { |
522 | 0 | constexpr int kNumRows = 10000; |
523 | 0 | TestLargeBackfill(kNumRows); |
524 | |
|
525 | 0 | const size_t num_tablets = cluster_->num_tablet_servers() * kTabletsPerServer; |
526 | 0 | const size_t min_expected_calls = static_cast<size_t>( |
527 | 0 | ceil(kNumRows / (kBackfillRpcDeadlineSmallMs * kBackfillRateRowsPerSec * 0.001))); |
528 | 0 | ASSERT_GT(min_expected_calls, num_tablets); |
529 | 0 | auto actual_calls = ASSERT_RESULT(TotalBackfillRpcCalls(cluster_.get())); |
530 | 0 | ASSERT_GE(actual_calls, num_tablets); |
531 | 0 | ASSERT_GE(actual_calls, min_expected_calls); |
532 | |
|
533 | 0 | auto avg_rpc_latency_usec = ASSERT_RESULT(AvgBackfillRpcLatencyInMicros(cluster_.get())); |
534 | 0 | LOG(INFO) << "Avg backfill latency was " << avg_rpc_latency_usec << " us"; |
535 | 0 | ASSERT_LE(avg_rpc_latency_usec, kBackfillRpcDeadlineSmallMs * 1000); |
536 | 0 | } |
537 | | |
538 | | // Make sure that CREATE INDEX NONCONCURRENTLY doesn't use backfill. |
539 | 0 | TEST_F(PgIndexBackfillTest, YB_DISABLE_TEST_IN_TSAN(Nonconcurrent)) { |
540 | 0 | auto client = ASSERT_RESULT(cluster_->CreateClient()); |
541 | |
|
542 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE TABLE $0 (i int)", kTableName)); |
543 | 0 | const std::string table_id = ASSERT_RESULT(GetTableIdByTableName( |
544 | 0 | client.get(), kDatabaseName, kTableName)); |
545 | | |
546 | | // To determine whether the index uses backfill or not, look at the table schema version before |
547 | | // and after. We can't look at the DocDB index permissions because |
548 | | // - if backfill is skipped, index_permissions is unset, and the default value is |
549 | | // INDEX_PERM_READ_WRITE_AND_DELETE |
550 | | // - if backfill is used, index_permissions is INDEX_PERM_READ_WRITE_AND_DELETE |
551 | | // - GetTableSchemaById offers no way to see whether the default value for index permissions is |
552 | | // set |
553 | 0 | std::shared_ptr<client::YBTableInfo> info = std::make_shared<client::YBTableInfo>(); |
554 | 0 | { |
555 | 0 | Synchronizer sync; |
556 | 0 | ASSERT_OK(client->GetTableSchemaById(table_id, info, sync.AsStatusCallback())); |
557 | 0 | ASSERT_OK(sync.Wait()); |
558 | 0 | } |
559 | 0 | ASSERT_EQ(info->schema.version(), 0); |
560 | |
|
561 | 0 | ASSERT_OK(conn_->ExecuteFormat( |
562 | 0 | "CREATE INDEX NONCONCURRENTLY $0 ON $1 (i)", |
563 | 0 | kIndexName, |
564 | 0 | kTableName)); |
565 | | |
566 | | // If the index used backfill, it would have incremented the table schema version by two or three: |
567 | | // - add index info with INDEX_PERM_DELETE_ONLY |
568 | | // - update to INDEX_PERM_DO_BACKFILL (as part of issue #6218) |
569 | | // - update to INDEX_PERM_READ_WRITE_AND_DELETE |
570 | | // If the index did not use backfill, it would have incremented the table schema version by one: |
571 | | // - add index info with no DocDB permission (default INDEX_PERM_READ_WRITE_AND_DELETE) |
572 | | // Expect that it did not use backfill. |
573 | 0 | { |
574 | 0 | Synchronizer sync; |
575 | 0 | ASSERT_OK(client->GetTableSchemaById(table_id, info, sync.AsStatusCallback())); |
576 | 0 | ASSERT_OK(sync.Wait()); |
577 | 0 | } |
578 | 0 | ASSERT_EQ(info->schema.version(), 1); |
579 | 0 | } |
580 | | |
581 | | class PgIndexBackfillTestSimultaneously : public PgIndexBackfillTest { |
582 | | public: |
583 | 0 | void UpdateMiniClusterOptions(ExternalMiniClusterOptions* options) override { |
584 | 0 | options->extra_tserver_flags.push_back( |
585 | 0 | Format("--ysql_pg_conf_csv=yb_index_state_flags_update_delay=$0", |
586 | 0 | kIndexStateFlagsUpdateDelay.ToMilliseconds())); |
587 | 0 | } |
588 | | protected: |
589 | | #ifdef NDEBUG // release build; see issue #6238 |
590 | | const MonoDelta kIndexStateFlagsUpdateDelay = 5s; |
591 | | #else // NDEBUG |
592 | | const MonoDelta kIndexStateFlagsUpdateDelay = 1s; |
593 | | #endif // NDEBUG |
594 | | }; |
595 | | |
596 | | // Test simultaneous CREATE INDEX. |
597 | | TEST_F_EX(PgIndexBackfillTest, |
598 | | YB_DISABLE_TEST_IN_TSAN(CreateIndexSimultaneously), |
599 | 0 | PgIndexBackfillTestSimultaneously) { |
600 | 0 | const std::string query = Format("SELECT * FROM $0 WHERE i = $1", kTableName, 7); |
601 | 0 | constexpr int kNumRows = 10; |
602 | 0 | constexpr int kNumThreads = 5; |
603 | 0 | int expected_schema_version = 0; |
604 | 0 | TestThreadHolder thread_holder; |
605 | |
|
606 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE TABLE $0 (i int)", kTableName)); |
607 | 0 | ASSERT_OK(conn_->ExecuteFormat( |
608 | 0 | "INSERT INTO $0 VALUES (generate_series(1, $1))", |
609 | 0 | kTableName, |
610 | 0 | kNumRows)); |
611 | |
|
612 | 0 | std::array<Status, kNumThreads> statuses; |
613 | 0 | for (int i = 0; i < kNumThreads; ++i) { |
614 | 0 | thread_holder.AddThreadFunctor([i, this, &statuses] { |
615 | 0 | LOG(INFO) << "Begin thread " << i; |
616 | 0 | PGConn create_conn = ASSERT_RESULT(ConnectToDB(kDatabaseName)); |
617 | 0 | statuses[i] = MoveStatus(create_conn.ExecuteFormat( |
618 | 0 | "CREATE INDEX $0 ON $1 (i)", |
619 | 0 | kIndexName, kTableName)); |
620 | 0 | }); |
621 | 0 | } |
622 | 0 | thread_holder.JoinAll(); |
623 | |
|
624 | 0 | LOG(INFO) << "Inspecting statuses"; |
625 | 0 | int num_ok = 0; |
626 | 0 | ASSERT_EQ(statuses.size(), kNumThreads); |
627 | 0 | for (const auto& status : statuses) { |
628 | 0 | if (status.ok()) { |
629 | 0 | num_ok++; |
630 | 0 | LOG(INFO) << "got ok status"; |
631 | | // Success index creations do two schema changes: |
632 | | // - add index with INDEX_PERM_WRITE_AND_DELETE |
633 | | // - transition to success INDEX_PERM_READ_WRITE_AND_DELETE |
634 | | // TODO(jason): change this when closing #6218 because DO_BACKFILL permission will add another |
635 | | // schema version. |
636 | 0 | expected_schema_version += 2; |
637 | 0 | } else { |
638 | 0 | ASSERT_TRUE(status.IsNetworkError()) << status; |
639 | 0 | const std::string msg = status.message().ToBuffer(); |
640 | 0 | const std::string relation_already_exists_msg = Format( |
641 | 0 | "relation \"$0\" already exists", kIndexName); |
642 | 0 | const std::vector<std::string> allowed_msgs{ |
643 | 0 | "Catalog Version Mismatch", |
644 | 0 | "Conflicts with higher priority transaction", |
645 | 0 | "Restart read required", |
646 | 0 | "Transaction aborted", |
647 | 0 | "Transaction metadata missing", |
648 | 0 | "Unknown transaction, could be recently aborted", |
649 | 0 | relation_already_exists_msg, |
650 | 0 | }; |
651 | 0 | ASSERT_TRUE(std::find_if( |
652 | 0 | std::begin(allowed_msgs), |
653 | 0 | std::end(allowed_msgs), |
654 | 0 | [&msg] (const std::string allowed_msg) { |
655 | 0 | return msg.find(allowed_msg) != std::string::npos; |
656 | 0 | }) != std::end(allowed_msgs)) |
657 | 0 | << status; |
658 | 0 | LOG(INFO) << "ignoring conflict error: " << status.message().ToBuffer(); |
659 | 0 | if (msg.find("Restart read required") == std::string::npos |
660 | 0 | && msg.find(relation_already_exists_msg) == std::string::npos) { |
661 | | // Failed index creations do two schema changes: |
662 | | // - add index with INDEX_PERM_WRITE_AND_DELETE |
663 | | // - remove index because of DDL transaction rollback ("Table transaction failed, deleting") |
664 | 0 | expected_schema_version += 2; |
665 | 0 | } else { |
666 | | // If the DocDB index was never created in the first place, it incurs no schema changes. |
667 | 0 | } |
668 | 0 | } |
669 | 0 | } |
670 | 0 | ASSERT_EQ(num_ok, 1) << "only one CREATE INDEX should succeed"; |
671 | | |
672 | 0 | LOG(INFO) << "Checking postgres schema"; |
673 | 0 | { |
674 | | // Check number of indexes. |
675 | 0 | PGResultPtr res = ASSERT_RESULT(conn_->FetchFormat( |
676 | 0 | "SELECT indexname FROM pg_indexes WHERE tablename = '$0'", kTableName)); |
677 | 0 | ASSERT_EQ(PQntuples(res.get()), 1); |
678 | 0 | const std::string actual = ASSERT_RESULT(GetString(res.get(), 0, 0)); |
679 | 0 | ASSERT_EQ(actual, kIndexName); |
680 | | |
681 | | // Check whether index is public using index scan. |
682 | 0 | ASSERT_TRUE(ASSERT_RESULT(conn_->HasIndexScan(query))); |
683 | 0 | } |
684 | 0 | LOG(INFO) << "Checking DocDB schema"; |
685 | 0 | std::vector<TableId> orphaned_docdb_index_ids; |
686 | 0 | { |
687 | 0 | auto client = ASSERT_RESULT(cluster_->CreateClient()); |
688 | 0 | const std::string table_id = ASSERT_RESULT(GetTableIdByTableName( |
689 | 0 | client.get(), kDatabaseName, kTableName)); |
690 | 0 | std::shared_ptr<client::YBTableInfo> table_info = std::make_shared<client::YBTableInfo>(); |
691 | 0 | Synchronizer sync; |
692 | 0 | ASSERT_OK(client->GetTableSchemaById(table_id, table_info, sync.AsStatusCallback())); |
693 | 0 | ASSERT_OK(sync.Wait()); |
694 | | |
695 | | // Check number of DocDB indexes. Normally, failed indexes should be cleaned up ("Table |
696 | | // transaction failed, deleting"), but in the event of an unexpected issue, they may not be. |
697 | | // (Not necessarily a fatal issue because the postgres schema is good.) |
698 | 0 | auto num_docdb_indexes = table_info->index_map.size(); |
699 | 0 | if (num_docdb_indexes > 1) { |
700 | 0 | LOG(INFO) << "found " << num_docdb_indexes << " DocDB indexes"; |
701 | | // These failed indexes not getting rolled back mean one less schema change each. Therefore, |
702 | | // adjust the expected schema version. |
703 | 0 | auto num_failed_docdb_indexes = num_docdb_indexes - 1; |
704 | 0 | expected_schema_version -= num_failed_docdb_indexes; |
705 | 0 | } |
706 | | |
707 | | // Check index permissions. Also collect orphaned DocDB indexes. |
708 | 0 | int num_rwd = 0; |
709 | 0 | for (const auto& pair : table_info->index_map) { |
710 | 0 | VLOG(1) << "table id: " << pair.first; |
711 | 0 | IndexPermissions perm = pair.second.index_permissions(); |
712 | 0 | if (perm == IndexPermissions::INDEX_PERM_READ_WRITE_AND_DELETE) { |
713 | 0 | num_rwd++; |
714 | 0 | } else { |
715 | 0 | ASSERT_EQ(perm, IndexPermissions::INDEX_PERM_WRITE_AND_DELETE); |
716 | 0 | orphaned_docdb_index_ids.emplace_back(pair.first); |
717 | 0 | } |
718 | 0 | } |
719 | 0 | ASSERT_EQ(num_rwd, 1) |
720 | 0 | << "found " << num_rwd << " fully created (readable) DocDB indexes: expected " << 1; |
721 | | |
722 | | // Check schema version. |
723 | 0 | ASSERT_EQ(table_info->schema.version(), expected_schema_version) |
724 | 0 | << "got indexed table schema version " << table_info->schema.version() |
725 | 0 | << ": expected " << expected_schema_version; |
726 | | // At least one index must have tried to create but gotten aborted, resulting in +1 or +2 |
727 | | // catalog version bump. The 2 below is for the successfully created index. |
728 | 0 | ASSERT_GT(expected_schema_version, 2); |
729 | 0 | } |
730 | |
|
731 | 0 | LOG(INFO) << "Checking if index still works"; |
732 | 0 | { |
733 | 0 | ASSERT_TRUE(ASSERT_RESULT(conn_->HasIndexScan(query))); |
734 | 0 | PGResultPtr res = ASSERT_RESULT(conn_->Fetch(query)); |
735 | 0 | ASSERT_EQ(PQntuples(res.get()), 1); |
736 | 0 | int32_t value = ASSERT_RESULT(GetInt32(res.get(), 0, 0)); |
737 | 0 | ASSERT_EQ(value, 7); |
738 | 0 | } |
739 | 0 | } |
740 | | |
741 | | // Make sure that backfill works in a tablegroup. |
742 | 0 | TEST_F(PgIndexBackfillTest, YB_DISABLE_TEST_IN_TSAN(Tablegroup)) { |
743 | 0 | const std::string kTablegroupName = "test_tgroup"; |
744 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE TABLEGROUP $0", kTablegroupName)); |
745 | |
|
746 | 0 | TestSimpleBackfill(Format("TABLEGROUP $0", kTablegroupName)); |
747 | 0 | } |
748 | | |
749 | | // Test that retain_delete_markers is properly set after index backfill. |
750 | 0 | TEST_F(PgIndexBackfillTest, YB_DISABLE_TEST_IN_TSAN(RetainDeleteMarkers)) { |
751 | 0 | TestRetainDeleteMarkers(kDatabaseName); |
752 | 0 | } |
753 | | |
754 | | // Override the index backfill test to do alter slowly. |
755 | | class PgIndexBackfillAlterSlowly : public PgIndexBackfillTest { |
756 | | public: |
757 | 0 | void UpdateMiniClusterOptions(ExternalMiniClusterOptions* options) override { |
758 | 0 | PgIndexBackfillTest::UpdateMiniClusterOptions(options); |
759 | 0 | options->extra_tserver_flags.push_back("--TEST_alter_schema_delay_ms=10000"); |
760 | 0 | } |
761 | | }; |
762 | | |
763 | | // Test whether IsCreateTableDone works when creating an index with backfill enabled. See issue |
764 | | // #6234. |
765 | | TEST_F_EX(PgIndexBackfillTest, |
766 | | YB_DISABLE_TEST_IN_TSAN(IsCreateTableDone), |
767 | 0 | PgIndexBackfillAlterSlowly) { |
768 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE TABLE $0 (i int)", kTableName)); |
769 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE INDEX ON $0 (i)", kTableName)); |
770 | 0 | } |
771 | | |
772 | | // Override the index backfill test to have different HBA config: |
773 | | // 1. if any user tries to access the authdb database, enforce md5 auth |
774 | | // 2. if the postgres user tries to access the yugabyte database, allow it |
775 | | // 3. if the yugabyte user tries to access the yugabyte database, allow it |
776 | | // 4. otherwise, disallow it |
777 | | class PgIndexBackfillAuth : public PgIndexBackfillTest { |
778 | | public: |
779 | 0 | void UpdateMiniClusterOptions(ExternalMiniClusterOptions* options) override { |
780 | 0 | PgIndexBackfillTest::UpdateMiniClusterOptions(options); |
781 | 0 | options->extra_tserver_flags.push_back(Format( |
782 | 0 | "--ysql_hba_conf=" |
783 | 0 | "host $0 all all md5," |
784 | 0 | "host $1 postgres all trust," |
785 | 0 | "host $1 yugabyte all trust", |
786 | 0 | kAuthDbName, |
787 | 0 | kDatabaseName)); |
788 | 0 | } |
789 | | |
790 | | const std::string kAuthDbName = "authdb"; |
791 | | }; |
792 | | |
793 | | // Test backfill on clusters where the yugabyte role has authentication enabled. |
794 | | TEST_F_EX(PgIndexBackfillTest, |
795 | | YB_DISABLE_TEST_IN_TSAN(Auth), |
796 | 0 | PgIndexBackfillAuth) { |
797 | 0 | LOG(INFO) << "create " << this->kAuthDbName << " database"; |
798 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE DATABASE $0", this->kAuthDbName)); |
799 | |
|
800 | 0 | LOG(INFO) << "backfill table on " << this->kAuthDbName << " database"; |
801 | 0 | { |
802 | 0 | const std::string& host = pg_ts->bind_host(); |
803 | 0 | const uint16_t port = pg_ts->pgsql_rpc_port(); |
804 | |
|
805 | 0 | PGConn auth_conn = ASSERT_RESULT(ConnectUsingString(Format( |
806 | 0 | "user=$0 password=$1 host=$2 port=$3 dbname=$4", |
807 | 0 | "yugabyte", |
808 | 0 | "yugabyte", |
809 | 0 | host, |
810 | 0 | port, |
811 | 0 | this->kAuthDbName))); |
812 | 0 | ASSERT_OK(auth_conn.ExecuteFormat("CREATE TABLE $0 (i int)", kTableName)); |
813 | 0 | ASSERT_OK(auth_conn.ExecuteFormat("CREATE INDEX ON $0 (i)", kTableName)); |
814 | 0 | } |
815 | 0 | } |
816 | | |
817 | | // Override the index backfill test to have HBA config with local trust: |
818 | | // 1. if any user tries to connect over ip, trust |
819 | | // 2. if any user tries to connect over unix-domain socket, trust |
820 | | class PgIndexBackfillLocalTrust : public PgIndexBackfillTest { |
821 | | public: |
822 | 0 | void UpdateMiniClusterOptions(ExternalMiniClusterOptions* options) override { |
823 | 0 | PgIndexBackfillTest::UpdateMiniClusterOptions(options); |
824 | 0 | options->extra_tserver_flags.push_back(Format( |
825 | 0 | "--ysql_hba_conf=" |
826 | 0 | "host $0 all all trust," |
827 | 0 | "local $0 all trust", |
828 | 0 | kDatabaseName)); |
829 | 0 | } |
830 | | }; |
831 | | |
832 | | // Make sure backfill works when there exists user-defined HBA configuration with "local". |
833 | | // This is for issue (#7705). |
834 | | TEST_F_EX(PgIndexBackfillTest, |
835 | | YB_DISABLE_TEST_IN_TSAN(LocalTrustSimple), |
836 | 0 | PgIndexBackfillLocalTrust) { |
837 | 0 | TestSimpleBackfill(); |
838 | 0 | } |
839 | | |
840 | | // Override the index backfill test to disable transparent retries on cache version mismatch. |
841 | | class PgIndexBackfillNoRetry : public PgIndexBackfillTest { |
842 | | public: |
843 | 0 | void UpdateMiniClusterOptions(ExternalMiniClusterOptions* options) override { |
844 | 0 | PgIndexBackfillTest::UpdateMiniClusterOptions(options); |
845 | 0 | options->extra_tserver_flags.push_back( |
846 | 0 | "--TEST_ysql_disable_transparent_cache_refresh_retry=true"); |
847 | 0 | } |
848 | | }; |
849 | | |
850 | | TEST_F_EX(PgIndexBackfillTest, |
851 | | YB_DISABLE_TEST_IN_TSAN(DropNoRetry), |
852 | 0 | PgIndexBackfillNoRetry) { |
853 | 0 | constexpr int kNumRows = 5; |
854 | |
|
855 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE TABLE $0 (i int, j int)", kTableName)); |
856 | 0 | ASSERT_OK(conn_->ExecuteFormat( |
857 | 0 | "INSERT INTO $0 VALUES (generate_series(1, $1), generate_series(11, 10 + $1))", |
858 | 0 | kTableName, |
859 | 0 | kNumRows)); |
860 | | |
861 | | // Create index. |
862 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE INDEX $0 ON $1 (i ASC)", kIndexName, kTableName)); |
863 | | |
864 | | // Update the table cache entry for the indexed table. |
865 | 0 | ASSERT_OK(conn_->FetchFormat("SELECT * FROM $0", kTableName)); |
866 | | |
867 | | // Drop index. |
868 | 0 | ASSERT_OK(conn_->ExecuteFormat("DROP INDEX $0", kIndexName)); |
869 | | |
870 | | // Ensure that there is no schema version mismatch for the indexed table. This is because the |
871 | | // above `DROP INDEX` should have invalidated the corresponding table cache entry. (There also |
872 | | // should be no catalog version mismatch because it is updated for the same session after DDL.) |
873 | 0 | ASSERT_OK(conn_->FetchFormat("SELECT * FROM $0", kTableName)); |
874 | 0 | } |
875 | | |
876 | | // Override the index backfill test to have delays for testing snapshot too old. |
877 | | class PgIndexBackfillSnapshotTooOld : public PgIndexBackfillTest { |
878 | | public: |
879 | 0 | void UpdateMiniClusterOptions(ExternalMiniClusterOptions* options) override { |
880 | 0 | PgIndexBackfillTest::UpdateMiniClusterOptions(options); |
881 | 0 | options->extra_tserver_flags.push_back("--TEST_slowdown_backfill_by_ms=10000"); |
882 | 0 | options->extra_tserver_flags.push_back( |
883 | 0 | "--ysql_pg_conf_csv=yb_index_state_flags_update_delay=0"); |
884 | 0 | options->extra_tserver_flags.push_back("--timestamp_history_retention_interval_sec=3"); |
885 | 0 | } |
886 | | }; |
887 | | |
888 | | // Make sure that index backfill doesn't care about snapshot too old. Force a situation where the |
889 | | // indexed table scan for backfill would occur after the committed history cutoff. A compaction is |
890 | | // needed to update this committed history cutoff, and the retention period needs to be low enough |
891 | | // so that the cutoff is ahead of backfill's safe read time. See issue #6333. |
892 | | TEST_F_EX(PgIndexBackfillTest, |
893 | | YB_DISABLE_TEST_IN_TSAN(SnapshotTooOld), |
894 | 0 | PgIndexBackfillSnapshotTooOld) { |
895 | 0 | auto client = ASSERT_RESULT(cluster_->CreateClient()); |
896 | 0 | constexpr int kTimeoutSec = 3; |
897 | 0 | TestThreadHolder thread_holder; |
898 | | |
899 | | // (Make it one tablet for simplicity.) |
900 | 0 | LOG(INFO) << "Create table..."; |
901 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE TABLE $0 (c char) SPLIT INTO 1 TABLETS", kTableName)); |
902 | |
|
903 | 0 | LOG(INFO) << "Get table id for indexed table..."; |
904 | 0 | const std::string table_id = ASSERT_RESULT(GetTableIdByTableName( |
905 | 0 | client.get(), kDatabaseName, kTableName)); |
906 | | |
907 | | // Insert something so that reading it would trigger snapshot too old. |
908 | 0 | ASSERT_OK(conn_->ExecuteFormat("INSERT INTO $0 VALUES ('s')", kTableName)); |
909 | | |
910 | | // conn_ should be used by at most one thread for thread safety. |
911 | 0 | thread_holder.AddThreadFunctor([this] { |
912 | 0 | LOG(INFO) << "Begin create thread"; |
913 | 0 | LOG(INFO) << "Create index..."; |
914 | 0 | Status s = conn_->ExecuteFormat("CREATE INDEX $0 ON $1 (c)", kIndexName, kTableName); |
915 | 0 | if (!s.ok()) { |
916 | | // We are doomed to fail the test. Before that, let's see if it turns out to be "snapshot too |
917 | | // old" or some other unexpected error. |
918 | 0 | ASSERT_TRUE(s.IsNetworkError()) << "got unexpected error: " << s; |
919 | 0 | ASSERT_TRUE(s.message().ToBuffer().find("Snapshot too old") != std::string::npos) |
920 | 0 | << "got unexpected error: " << s; |
921 | | // It is "snapshot too old". Fail now. |
922 | 0 | FAIL() << "got snapshot too old: " << s; |
923 | 0 | } |
924 | 0 | }); |
925 | 0 | thread_holder.AddThreadFunctor([&client, &table_id] { |
926 | 0 | LOG(INFO) << "Begin compact thread"; |
927 | | // Sleep until we are in the interval |
928 | | // (read_time + history_retention_interval, read_time + slowdown_backfill) |
929 | | // = (read_time + 3s, read_time + 10s) |
930 | | // Choose read_time + 5s. |
931 | 0 | LOG(INFO) << "Sleep..."; |
932 | 0 | SleepFor(1s); // approximate setup time before getting to the backfill stage |
933 | 0 | SleepFor(5s); |
934 | |
|
935 | 0 | LOG(INFO) << "Flush and compact indexed table..."; |
936 | 0 | ASSERT_OK(client->FlushTables( |
937 | 0 | {table_id}, |
938 | 0 | false /* add_indexes */, |
939 | 0 | kTimeoutSec, |
940 | 0 | false /* is_compaction */)); |
941 | 0 | ASSERT_OK(client->FlushTables( |
942 | 0 | {table_id}, |
943 | 0 | false /* add_indexes */, |
944 | 0 | kTimeoutSec, |
945 | 0 | true /* is_compaction */)); |
946 | 0 | }); |
947 | 0 | thread_holder.JoinAll(); |
948 | 0 | } |
949 | | |
950 | | // Override the index backfill test to have slower backfill-related operations |
951 | | class PgIndexBackfillSlow : public PgIndexBackfillTest { |
952 | | public: |
953 | 0 | void UpdateMiniClusterOptions(ExternalMiniClusterOptions* options) override { |
954 | 0 | PgIndexBackfillTest::UpdateMiniClusterOptions(options); |
955 | 0 | options->extra_master_flags.push_back(Format( |
956 | 0 | "--TEST_slowdown_backfill_alter_table_rpcs_ms=$0", |
957 | 0 | kBackfillAlterTableDelay.ToMilliseconds())); |
958 | 0 | options->extra_tserver_flags.push_back(Format( |
959 | 0 | "--ysql_pg_conf_csv=yb_index_state_flags_update_delay=$0", |
960 | 0 | kIndexStateFlagsUpdateDelay.ToMilliseconds())); |
961 | 0 | options->extra_tserver_flags.push_back(Format( |
962 | 0 | "--TEST_slowdown_backfill_by_ms=$0", |
963 | 0 | kBackfillDelay.ToMilliseconds())); |
964 | 0 | } |
965 | | |
966 | | protected: |
967 | | Result<bool> IsAtTargetIndexStateFlags( |
968 | | const std::string& index_name, |
969 | 0 | const IndexStateFlags& target_index_state_flags) { |
970 | 0 | Result<IndexStateFlags> res = GetIndexStateFlags(index_name); |
971 | 0 | IndexStateFlags actual_index_state_flags; |
972 | 0 | if (res.ok()) { |
973 | 0 | actual_index_state_flags = res.get(); |
974 | 0 | } else if (res.status().IsNotFound()) { |
975 | 0 | LOG(WARNING) << res.status(); |
976 | 0 | return false; |
977 | 0 | } else { |
978 | 0 | return res.status(); |
979 | 0 | } |
980 | | |
981 | 0 | if (actual_index_state_flags < target_index_state_flags) { |
982 | 0 | LOG(INFO) << index_name |
983 | 0 | << " not yet at target index state flags " |
984 | 0 | << ToString(target_index_state_flags); |
985 | 0 | return false; |
986 | 0 | } else if (actual_index_state_flags > target_index_state_flags) { |
987 | 0 | return STATUS(RuntimeError, |
988 | 0 | Format("$0 exceeded target index state flags $1", |
989 | 0 | index_name, |
990 | 0 | target_index_state_flags)); |
991 | 0 | } |
992 | 0 | return true; |
993 | 0 | } |
994 | | |
995 | | CHECKED_STATUS WaitForBackfillSafeTime( |
996 | 0 | const client::YBTableName& table_name, const std::string& index_name) { |
997 | 0 | LOG(INFO) << "Waiting for pg_index indislive to be true"; |
998 | 0 | RETURN_NOT_OK(WaitFor( |
999 | 0 | [this, &index_name] { |
1000 | 0 | return IsAtTargetIndexStateFlags(index_name, IndexStateFlags{IndexStateFlag::kIndIsLive}); |
1001 | 0 | }, |
1002 | 0 | kCreateIndexStartupGracePeriod, |
1003 | 0 | "Wait for pg_index indislive=true", |
1004 | 0 | MonoDelta::FromMilliseconds(test_util::kDefaultInitialWaitMs), |
1005 | 0 | test_util::kDefaultWaitDelayMultiplier, |
1006 | 0 | kMaxDelay)); |
1007 | |
|
1008 | 0 | LOG(INFO) << "Waiting for pg_index indisready to be true"; |
1009 | 0 | RETURN_NOT_OK(WaitFor( |
1010 | 0 | [this, &index_name] { |
1011 | 0 | return IsAtTargetIndexStateFlags( |
1012 | 0 | index_name, IndexStateFlags{IndexStateFlag::kIndIsLive, IndexStateFlag::kIndIsReady}); |
1013 | 0 | }, |
1014 | 0 | kIndexStateFlagsUpdateDelay + kIndexStateFlagsUpdateGracePeriod, |
1015 | 0 | "Wait for pg_index indisready=true", |
1016 | 0 | kIndexStateFlagsUpdateDelay - kMaxDelay /* initial_delay */, |
1017 | 0 | test_util::kDefaultWaitDelayMultiplier, |
1018 | 0 | kMaxDelay)); |
1019 | |
|
1020 | 0 | LOG(INFO) << "Waiting till (approx) the end of the delay after committing indisready true"; |
1021 | 0 | SleepFor(kIndexStateFlagsUpdateDelay); |
1022 | |
|
1023 | 0 | auto client = VERIFY_RESULT(cluster_->CreateClient()); |
1024 | 0 | const std::string table_id = VERIFY_RESULT( |
1025 | 0 | GetTableIdByTableName(client.get(), table_name.namespace_name(), table_name.table_name())); |
1026 | 0 | RETURN_NOT_OK(WaitForBackfillSafeTimeOn( |
1027 | 0 | cluster_->GetLeaderMasterProxy<master::MasterDdlProxy>(), table_id)); |
1028 | |
|
1029 | 0 | return Status::OK(); |
1030 | 0 | } |
1031 | | |
1032 | | // buffer times for unexpected situations. |
1033 | | const MonoDelta kCreateIndexStartupGracePeriod = 30s; |
1034 | | const MonoDelta kIndexStateFlagsUpdateGracePeriod = 5s; |
1035 | | |
1036 | | // gflag delay times. |
1037 | | const MonoDelta kBackfillAlterTableDelay = 0s; |
1038 | | const MonoDelta kBackfillDelay = RegularBuildVsSanitizers(3s, 7s); |
1039 | | const MonoDelta kIndexStateFlagsUpdateDelay = RegularBuildVsSanitizers(3s, 7s); |
1040 | | |
1041 | | // maximum delay between checks on index state flags. |
1042 | | const MonoDelta kMaxDelay = 100ms; |
1043 | | |
1044 | | TestThreadHolder thread_holder_; |
1045 | | |
1046 | | private: |
1047 | 0 | Result<IndexStateFlags> GetIndexStateFlags(const std::string& index_name) { |
1048 | 0 | const std::string quoted_index_name = PqEscapeLiteral(index_name); |
1049 | |
|
1050 | 0 | PGResultPtr res = VERIFY_RESULT(conn_->FetchFormat( |
1051 | 0 | "SELECT indislive, indisready, indisvalid" |
1052 | 0 | " FROM pg_class INNER JOIN pg_index ON pg_class.oid = pg_index.indexrelid" |
1053 | 0 | " WHERE pg_class.relname = $0", |
1054 | 0 | quoted_index_name)); |
1055 | 0 | if (PQntuples(res.get()) == 0) { |
1056 | 0 | return STATUS_FORMAT(NotFound, "$0 not found in pg_class and/or pg_index", quoted_index_name); |
1057 | 0 | } |
1058 | 0 | if (int num_cols = PQnfields(res.get()) != 3) { |
1059 | 0 | return STATUS_FORMAT(Corruption, "got unexpected number of columns: $0", num_cols); |
1060 | 0 | } |
1061 | | |
1062 | 0 | IndexStateFlags index_state_flags; |
1063 | 0 | if (VERIFY_RESULT(GetBool(res.get(), 0, 0))) { |
1064 | 0 | index_state_flags.Set(IndexStateFlag::kIndIsLive); |
1065 | 0 | } |
1066 | 0 | if (VERIFY_RESULT(GetBool(res.get(), 0, 1))) { |
1067 | 0 | index_state_flags.Set(IndexStateFlag::kIndIsReady); |
1068 | 0 | } |
1069 | 0 | if (VERIFY_RESULT(GetBool(res.get(), 0, 2))) { |
1070 | 0 | index_state_flags.Set(IndexStateFlag::kIndIsValid); |
1071 | 0 | } |
1072 | |
|
1073 | 0 | return index_state_flags; |
1074 | 0 | } |
1075 | | }; |
1076 | | |
1077 | | // Make sure that read time (and write time) for backfill works. Simulate the following: |
1078 | | // Session A Session B |
1079 | | // -------------------------- --------------------------------- |
1080 | | // CREATE INDEX |
1081 | | // - indislive |
1082 | | // - indisready |
1083 | | // - backfill |
1084 | | // - get safe time for read |
1085 | | // UPDATE a row of the indexed table |
1086 | | // - do the actual backfill |
1087 | | // - indisvalid |
1088 | | // The backfill should use the values before update when writing to the index. The update should |
1089 | | // write and delete to the index because of permissions. Since backfill writes with an ancient |
1090 | | // timestamp, the update should appear to have happened after the backfill. |
1091 | | TEST_F_EX(PgIndexBackfillTest, |
1092 | | YB_DISABLE_TEST_IN_TSAN(ReadTime), |
1093 | 0 | PgIndexBackfillSlow) { |
1094 | 0 | ASSERT_OK(conn_->ExecuteFormat( |
1095 | 0 | "CREATE TABLE $0 (i int, j int, PRIMARY KEY (i ASC))", kTableName)); |
1096 | 0 | ASSERT_OK(conn_->ExecuteFormat( |
1097 | 0 | "INSERT INTO $0 VALUES (generate_series(0, 5), generate_series(10, 15))", kTableName)); |
1098 | | |
1099 | | // conn_ should be used by at most one thread for thread safety. |
1100 | 0 | thread_holder_.AddThreadFunctor([this] { |
1101 | 0 | LOG(INFO) << "Begin create thread"; |
1102 | 0 | PGConn create_conn = ASSERT_RESULT(ConnectToDB(kDatabaseName)); |
1103 | 0 | ASSERT_OK(create_conn.ExecuteFormat("CREATE INDEX $0 ON $1 (j ASC)", kIndexName, kTableName)); |
1104 | 0 | }); |
1105 | 0 | thread_holder_.AddThreadFunctor([this] { |
1106 | 0 | LOG(INFO) << "Begin write thread"; |
1107 | 0 | ASSERT_OK(WaitForBackfillSafeTime(kYBTableName, kIndexName)); |
1108 | |
|
1109 | 0 | LOG(INFO) << "Updating row"; |
1110 | 0 | ASSERT_OK(conn_->ExecuteFormat("UPDATE $0 SET j = j + 100 WHERE i = 3", kTableName)); |
1111 | 0 | LOG(INFO) << "Done updating row"; |
1112 | | |
1113 | | // It should still be in the backfill stage, hopefully before the actual backfill started. |
1114 | 0 | ASSERT_TRUE(ASSERT_RESULT(IsAtTargetIndexStateFlags( |
1115 | 0 | kIndexName, IndexStateFlags{IndexStateFlag::kIndIsLive, IndexStateFlag::kIndIsReady}))); |
1116 | 0 | }); |
1117 | 0 | thread_holder_.JoinAll(); |
1118 | | |
1119 | | // Index scan to verify contents of index table. |
1120 | 0 | const std::string query = Format("SELECT * FROM $0 WHERE j = 113", kTableName); |
1121 | 0 | ASSERT_OK(WaitFor( |
1122 | 0 | [this, &query] { |
1123 | 0 | return conn_->HasIndexScan(query); |
1124 | 0 | }, |
1125 | 0 | kIndexStateFlagsUpdateGracePeriod + kIndexStateFlagsUpdateDelay, |
1126 | 0 | "Wait for IndexScan")); |
1127 | 0 | PGResultPtr res = ASSERT_RESULT(conn_->Fetch(query)); |
1128 | 0 | int lines = PQntuples(res.get()); |
1129 | 0 | ASSERT_EQ(1, lines); |
1130 | 0 | int columns = PQnfields(res.get()); |
1131 | 0 | ASSERT_EQ(2, columns); |
1132 | 0 | int32_t key = ASSERT_RESULT(GetInt32(res.get(), 0, 0)); |
1133 | 0 | ASSERT_EQ(key, 3); |
1134 | | // Make sure that the update is visible. |
1135 | 0 | int32_t value = ASSERT_RESULT(GetInt32(res.get(), 0, 1)); |
1136 | 0 | ASSERT_EQ(value, 113); |
1137 | 0 | } |
1138 | | |
1139 | | // Make sure that updates at each stage of multi-stage CREATE INDEX work. Simulate the following: |
1140 | | // Session A Session B |
1141 | | // -------------------------- --------------------------------- |
1142 | | // CREATE INDEX |
1143 | | // - indislive |
1144 | | // UPDATE a row of the indexed table |
1145 | | // - indisready |
1146 | | // UPDATE a row of the indexed table |
1147 | | // - indisvalid |
1148 | | // UPDATE a row of the indexed table |
1149 | | // Updates should succeed and get written to the index. |
1150 | | TEST_F_EX(PgIndexBackfillTest, |
1151 | | YB_DISABLE_TEST_IN_TSAN(Permissions), |
1152 | 0 | PgIndexBackfillSlow) { |
1153 | 0 | const CoarseDuration kThreadWaitTime = 60s; |
1154 | 0 | const std::array<std::pair<IndexStateFlags, int>, 3> index_state_flags_key_pairs = { |
1155 | 0 | std::make_pair(IndexStateFlags{IndexStateFlag::kIndIsLive}, 2), |
1156 | 0 | std::make_pair(IndexStateFlags{IndexStateFlag::kIndIsLive, IndexStateFlag::kIndIsReady}, 3), |
1157 | 0 | std::make_pair(IndexStateFlags{ |
1158 | 0 | IndexStateFlag::kIndIsLive, |
1159 | 0 | IndexStateFlag::kIndIsReady, |
1160 | 0 | IndexStateFlag::kIndIsValid, |
1161 | 0 | }, 4), |
1162 | 0 | }; |
1163 | 0 | std::atomic<int> updates(0); |
1164 | |
|
1165 | 0 | ASSERT_OK(conn_->ExecuteFormat( |
1166 | 0 | "CREATE TABLE $0 (i int, j int, PRIMARY KEY (i ASC))", kTableName)); |
1167 | 0 | ASSERT_OK(conn_->ExecuteFormat( |
1168 | 0 | "INSERT INTO $0 VALUES (generate_series(0, 5), generate_series(10, 15))", kTableName)); |
1169 | | |
1170 | | // conn_ should be used by at most one thread for thread safety. |
1171 | 0 | thread_holder_.AddThreadFunctor([this] { |
1172 | 0 | LOG(INFO) << "Begin create thread"; |
1173 | 0 | PGConn create_conn = ASSERT_RESULT(ConnectToDB(kDatabaseName)); |
1174 | 0 | ASSERT_OK(create_conn.ExecuteFormat("CREATE INDEX $0 ON $1 (j ASC)", kIndexName, kTableName)); |
1175 | 0 | }); |
1176 | 0 | thread_holder_.AddThreadFunctor([this, &index_state_flags_key_pairs, &updates] { |
1177 | 0 | LOG(INFO) << "Begin write thread"; |
1178 | 0 | for (const auto& pair : index_state_flags_key_pairs) { |
1179 | 0 | const IndexStateFlags& index_state_flags = pair.first; |
1180 | 0 | int key = pair.second; |
1181 | |
|
1182 | 0 | MonoDelta timeout; |
1183 | 0 | if (index_state_flags == IndexStateFlags{IndexStateFlag::kIndIsLive}) { |
1184 | 0 | timeout = kCreateIndexStartupGracePeriod; |
1185 | 0 | } else if (index_state_flags == IndexStateFlags{ |
1186 | 0 | IndexStateFlag::kIndIsLive, |
1187 | 0 | IndexStateFlag::kIndIsReady, |
1188 | 0 | }) { |
1189 | 0 | timeout = kIndexStateFlagsUpdateDelay + kIndexStateFlagsUpdateGracePeriod; |
1190 | 0 | } else { |
1191 | 0 | ASSERT_TRUE((index_state_flags == IndexStateFlags{ |
1192 | 0 | IndexStateFlag::kIndIsLive, |
1193 | 0 | IndexStateFlag::kIndIsReady, |
1194 | 0 | IndexStateFlag::kIndIsValid, |
1195 | 0 | })); |
1196 | 0 | timeout = ((kIndexStateFlagsUpdateDelay + kIndexStateFlagsUpdateGracePeriod) |
1197 | 0 | + (kBackfillAlterTableDelay * 2) |
1198 | 0 | + (kBackfillDelay + kIndexStateFlagsUpdateGracePeriod)); |
1199 | 0 | } |
1200 | 0 | ASSERT_OK(WaitFor( |
1201 | 0 | [this, &index_state_flags] { |
1202 | 0 | return IsAtTargetIndexStateFlags(kIndexName, index_state_flags); |
1203 | 0 | }, |
1204 | 0 | timeout, |
1205 | 0 | Format("get index state flags: $0", index_state_flags))); |
1206 | 0 | LOG(INFO) << "running UPDATE on i = " << key; |
1207 | 0 | ASSERT_OK(conn_->ExecuteFormat("UPDATE $0 SET j = j + 100 WHERE i = $1", kTableName, key)); |
1208 | 0 | LOG(INFO) << "done running UPDATE on i = " << key; |
1209 | | |
1210 | | // Make sure permission didn't change yet. |
1211 | 0 | ASSERT_TRUE(ASSERT_RESULT(IsAtTargetIndexStateFlags(kIndexName, index_state_flags))); |
1212 | 0 | updates++; |
1213 | 0 | } |
1214 | 0 | }); |
1215 | 0 | thread_holder_.WaitAndStop(kThreadWaitTime); |
1216 | |
|
1217 | 0 | ASSERT_EQ(updates.load(std::memory_order_acquire), index_state_flags_key_pairs.size()); |
1218 | |
|
1219 | 0 | for (const auto& pair : index_state_flags_key_pairs) { |
1220 | 0 | int key = pair.second; |
1221 | | |
1222 | | // Verify contents of index table. |
1223 | 0 | const std::string query = Format( |
1224 | 0 | "WITH j_idx AS (SELECT * FROM $0 ORDER BY j) SELECT j FROM j_idx WHERE i = $1", |
1225 | 0 | kTableName, |
1226 | 0 | key); |
1227 | 0 | ASSERT_OK(WaitFor( |
1228 | 0 | [this, &query] { |
1229 | 0 | return conn_->HasIndexScan(query); |
1230 | 0 | }, |
1231 | 0 | kIndexStateFlagsUpdateGracePeriod + kIndexStateFlagsUpdateDelay, |
1232 | 0 | "Wait for IndexScan")); |
1233 | 0 | PGResultPtr res = ASSERT_RESULT(conn_->Fetch(query)); |
1234 | 0 | int lines = PQntuples(res.get()); |
1235 | 0 | ASSERT_EQ(1, lines); |
1236 | 0 | int columns = PQnfields(res.get()); |
1237 | 0 | ASSERT_EQ(1, columns); |
1238 | | // Make sure that the update is visible. |
1239 | 0 | int value = ASSERT_RESULT(GetInt32(res.get(), 0, 0)); |
1240 | 0 | ASSERT_EQ(value, key + 110); |
1241 | 0 | } |
1242 | 0 | } |
1243 | | |
1244 | | // Make sure that writes during CREATE UNIQUE INDEX don't cause unique duplicate row errors to be |
1245 | | // thrown. Simulate the following: |
1246 | | // Session A Session B |
1247 | | // -------------------------- --------------------------------- |
1248 | | // INSERT a row to the indexed table |
1249 | | // CREATE UNIQUE INDEX |
1250 | | // INSERT a row to the indexed table |
1251 | | // - indislive |
1252 | | // INSERT a row to the indexed table |
1253 | | // - indisready |
1254 | | // INSERT a row to the indexed table |
1255 | | // - backfill |
1256 | | // INSERT a row to the indexed table |
1257 | | // - indisvalid |
1258 | | // INSERT a row to the indexed table |
1259 | | // Particularly pay attention to the insert between indisready and backfill. The insert |
1260 | | // should cause a write to go to the index. Backfill should choose a read time after this write, so |
1261 | | // it should try to backfill this same row. Rather than conflicting when we see the row already |
1262 | | // exists in the index during backfill, check whether the rows match, and don't error if they do. |
1263 | | TEST_F_EX(PgIndexBackfillTest, |
1264 | | YB_DISABLE_TEST_IN_TSAN(CreateUniqueIndexWithOnlineWrites), |
1265 | 0 | PgIndexBackfillSlow) { |
1266 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE TABLE $0 (i int)", kTableName)); |
1267 | | |
1268 | | // Start a thread that continuously inserts distinct values. The hope is that this would cause |
1269 | | // inserts to happen at all permissions. |
1270 | 0 | thread_holder_.AddThreadFunctor([this, &stop = thread_holder_.stop_flag()] { |
1271 | 0 | LOG(INFO) << "Begin write thread"; |
1272 | 0 | PGConn insert_conn = ASSERT_RESULT(Connect()); |
1273 | 0 | int i = 0; |
1274 | 0 | while (!stop.load(std::memory_order_acquire)) { |
1275 | 0 | Status status = insert_conn.ExecuteFormat("INSERT INTO $0 VALUES ($1)", kTableName, ++i); |
1276 | 0 | if (!status.ok()) { |
1277 | | // Ignore transient errors that likely occur when changing index permissions. |
1278 | | // TODO(jason): no longer expect schema version mismatch errors after closing issue #3979. |
1279 | 0 | ASSERT_TRUE(status.IsNetworkError()) << status; |
1280 | 0 | std::string msg = status.message().ToBuffer(); |
1281 | 0 | const std::vector<std::string> allowed_msgs{ |
1282 | 0 | "Errors occurred while reaching out to the tablet servers", |
1283 | 0 | "Resource unavailable : RocksDB", |
1284 | 0 | "schema version mismatch", |
1285 | 0 | "Transaction aborted", |
1286 | 0 | "expired or aborted by a conflict", |
1287 | 0 | "Transaction was recently aborted", |
1288 | 0 | }; |
1289 | 0 | ASSERT_TRUE(std::find_if( |
1290 | 0 | std::begin(allowed_msgs), |
1291 | 0 | std::end(allowed_msgs), |
1292 | 0 | [&msg] (const std::string allowed_msg) { |
1293 | 0 | return msg.find(allowed_msg) != std::string::npos; |
1294 | 0 | }) != std::end(allowed_msgs)) |
1295 | 0 | << status; |
1296 | 0 | LOG(WARNING) << "ignoring transient error: " << status.message().ToBuffer(); |
1297 | 0 | } |
1298 | 0 | } |
1299 | 0 | }); |
1300 | | |
1301 | | // Create unique index (should not complain about duplicate row). |
1302 | 0 | LOG(INFO) << "Create unique index..."; |
1303 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE UNIQUE INDEX ON $0 (i ASC)", kTableName)); |
1304 | |
|
1305 | 0 | thread_holder_.Stop(); |
1306 | 0 | } |
1307 | | |
1308 | | // Simulate the following: |
1309 | | // Session A Session B |
1310 | | // ------------------------------------ ------------------------------------------- |
1311 | | // CREATE TABLE (i, j, PRIMARY KEY (i)) |
1312 | | // INSERT (1, 'a') |
1313 | | // CREATE UNIQUE INDEX (j) |
1314 | | // - DELETE_ONLY perm |
1315 | | // DELETE (1, 'a') |
1316 | | // (delete (1, 'a') to index) |
1317 | | // INSERT (2, 'a') |
1318 | | // - WRITE_DELETE perm |
1319 | | // - BACKFILL perm |
1320 | | // - get safe time for read |
1321 | | // INSERT (3, 'a') |
1322 | | // (insert (3, 'a') to index) |
1323 | | // - do the actual backfill |
1324 | | // (insert (2, 'a') to index--detect conflict) |
1325 | | // - READ_WRITE_DELETE perm |
1326 | | // This test is for issue #6208. |
1327 | | TEST_F_EX(PgIndexBackfillTest, |
1328 | | YB_DISABLE_TEST_IN_TSAN(CreateUniqueIndexWriteAfterSafeTime), |
1329 | 0 | PgIndexBackfillSlow) { |
1330 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE TABLE $0 (i int, j char, PRIMARY KEY (i))", kTableName)); |
1331 | 0 | ASSERT_OK(conn_->ExecuteFormat("INSERT INTO $0 VALUES (1, 'a')", kTableName)); |
1332 | | |
1333 | | // conn_ should be used by at most one thread for thread safety. |
1334 | 0 | thread_holder_.AddThreadFunctor([this] { |
1335 | 0 | LOG(INFO) << "Begin create thread"; |
1336 | 0 | LOG(INFO) << "Creating index..."; |
1337 | 0 | PGConn create_conn = ASSERT_RESULT(ConnectToDB(kDatabaseName)); |
1338 | 0 | Status s = create_conn.ExecuteFormat( |
1339 | 0 | "CREATE UNIQUE INDEX $0 ON $1 (j ASC)", kIndexName, kTableName); |
1340 | 0 | ASSERT_NOK(s); |
1341 | 0 | ASSERT_TRUE(s.IsNetworkError()); |
1342 | 0 | ASSERT_TRUE(s.message().ToBuffer().find("duplicate key value") != std::string::npos) << s; |
1343 | 0 | }); |
1344 | 0 | thread_holder_.AddThreadFunctor([this] { |
1345 | 0 | LOG(INFO) << "Begin write thread"; |
1346 | 0 | { |
1347 | 0 | const IndexStateFlags index_state_flags{IndexStateFlag::kIndIsLive}; |
1348 | |
|
1349 | 0 | LOG(INFO) << "Wait for indislive index state flag"; |
1350 | 0 | ASSERT_OK(WaitFor( |
1351 | 0 | [this, &index_state_flags] { |
1352 | 0 | return IsAtTargetIndexStateFlags(kIndexName, index_state_flags); |
1353 | 0 | }, |
1354 | 0 | kCreateIndexStartupGracePeriod, |
1355 | 0 | Format("get index state flags: $0", index_state_flags))); |
1356 | |
|
1357 | 0 | LOG(INFO) << "Do delete and insert"; |
1358 | 0 | ASSERT_OK(conn_->ExecuteFormat("DELETE FROM $0 WHERE i = 1", kTableName)); |
1359 | 0 | ASSERT_OK(conn_->ExecuteFormat("INSERT INTO $0 VALUES (2, 'a')", kTableName)); |
1360 | |
|
1361 | 0 | LOG(INFO) << "Check we're not yet at indisready index state flag"; |
1362 | 0 | ASSERT_TRUE(ASSERT_RESULT(IsAtTargetIndexStateFlags(kIndexName, index_state_flags))); |
1363 | 0 | } |
1364 | |
|
1365 | 0 | { |
1366 | 0 | const IndexStateFlags index_state_flags{ |
1367 | 0 | IndexStateFlag::kIndIsLive, |
1368 | 0 | IndexStateFlag::kIndIsReady, |
1369 | 0 | }; |
1370 | |
|
1371 | 0 | LOG(INFO) << "Wait for indisready index state flag"; |
1372 | 0 | ASSERT_OK(WaitFor( |
1373 | 0 | [this, &index_state_flags] { |
1374 | 0 | return IsAtTargetIndexStateFlags(kIndexName, index_state_flags); |
1375 | 0 | }, |
1376 | 0 | kIndexStateFlagsUpdateDelay + kIndexStateFlagsUpdateGracePeriod, |
1377 | 0 | Format("get index state flags: $0", index_state_flags))); |
1378 | 0 | } |
1379 | |
|
1380 | 0 | { |
1381 | 0 | LOG(INFO) << "Wait for backfill (approx)"; |
1382 | 0 | SleepFor(kIndexStateFlagsUpdateDelay); |
1383 | |
|
1384 | 0 | LOG(INFO) << "Wait to get safe time for backfill (approx)"; |
1385 | 0 | SleepFor(kBackfillDelay / 2); |
1386 | |
|
1387 | 0 | LOG(INFO) << "Do insert"; |
1388 | 0 | CoarseBackoffWaiter waiter( |
1389 | 0 | CoarseMonoClock::Now() + (kBackfillDelay / 2 + kIndexStateFlagsUpdateGracePeriod), |
1390 | 0 | CoarseMonoClock::Duration::max()); |
1391 | 0 | while (true) { |
1392 | 0 | Status status = conn_->ExecuteFormat("INSERT INTO $0 VALUES (3, 'a')", kTableName); |
1393 | 0 | LOG(INFO) << "Got " << yb::ToString(status); |
1394 | 0 | if (status.ok()) { |
1395 | 0 | break; |
1396 | 0 | } else { |
1397 | 0 | ASSERT_FALSE(status.IsIllegalState() && |
1398 | 0 | status.message().ToBuffer().find("Duplicate value") != std::string::npos) |
1399 | 0 | << "The insert should come before backfill, so it should not cause duplicate" |
1400 | 0 | << " conflict."; |
1401 | 0 | ASSERT_TRUE(waiter.Wait()); |
1402 | 0 | } |
1403 | 0 | } |
1404 | 0 | } |
1405 | 0 | }); |
1406 | 0 | thread_holder_.JoinAll(); |
1407 | | |
1408 | | // Check. |
1409 | 0 | { |
1410 | 0 | CoarseBackoffWaiter waiter(CoarseMonoClock::Now() + 10s, CoarseMonoClock::Duration::max()); |
1411 | 0 | while (true) { |
1412 | 0 | Result<PGResultPtr> result = conn_->FetchFormat("SELECT count(*) FROM $0", kTableName); |
1413 | 0 | if (result.ok()) { |
1414 | 0 | PGResultPtr res = std::move(*result); |
1415 | 0 | const int64_t main_table_size = ASSERT_RESULT(GetInt64(res.get(), 0, 0)); |
1416 | 0 | ASSERT_EQ(main_table_size, 2); |
1417 | 0 | break; |
1418 | 0 | } |
1419 | 0 | ASSERT_TRUE(result.status().IsQLError()) << result.status(); |
1420 | 0 | ASSERT_TRUE(result.status().message().ToBuffer().find("schema version mismatch") |
1421 | 0 | != std::string::npos) << result.status(); |
1422 | 0 | ASSERT_TRUE(waiter.Wait()); |
1423 | 0 | } |
1424 | 0 | } |
1425 | 0 | } |
1426 | | |
1427 | | // Simulate the following: |
1428 | | // Session A Session B |
1429 | | // ------------------------------------ ------------------------------------------- |
1430 | | // CREATE TABLE (i, j, PRIMARY KEY (i)) |
1431 | | // INSERT (1, 'a') |
1432 | | // CREATE UNIQUE INDEX (j) |
1433 | | // - indislive |
1434 | | // - indisready |
1435 | | // - backfill stage |
1436 | | // - get safe time for read |
1437 | | // DELETE (1, 'a') |
1438 | | // (delete (1, 'a') to index) |
1439 | | // - do the actual backfill |
1440 | | // (insert (1, 'a') to index) |
1441 | | // - indisvalid |
1442 | | // This test is for issue #6811. Remember, backfilled rows get written with write time = safe time, |
1443 | | // so they should have an MVCC timestamp lower than that of the deletion. If deletes to the index |
1444 | | // aren't written, then this test will always fail because the backfilled row has no delete to cover |
1445 | | // it. If deletes to the index aren't retained, then this test will fail if compactions get rid of |
1446 | | // the delete before the backfilled row gets written. |
1447 | | TEST_F_EX(PgIndexBackfillTest, |
1448 | | YB_DISABLE_TEST_IN_TSAN(RetainDeletes), |
1449 | 0 | PgIndexBackfillSlow) { |
1450 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE TABLE $0 (i int, j char, PRIMARY KEY (i))", kTableName)); |
1451 | 0 | ASSERT_OK(conn_->ExecuteFormat("INSERT INTO $0 VALUES (1, 'a')", kTableName)); |
1452 | | |
1453 | | // conn_ should be used by at most one thread for thread safety. |
1454 | 0 | thread_holder_.AddThreadFunctor([this] { |
1455 | 0 | LOG(INFO) << "Begin create thread"; |
1456 | 0 | LOG(INFO) << "Creating index"; |
1457 | 0 | PGConn create_conn = ASSERT_RESULT(ConnectToDB(kDatabaseName)); |
1458 | 0 | ASSERT_OK(create_conn.ExecuteFormat( |
1459 | 0 | "CREATE UNIQUE INDEX $0 ON $1 (j ASC)", kIndexName, kTableName)); |
1460 | 0 | }); |
1461 | 0 | thread_holder_.AddThreadFunctor([this] { |
1462 | 0 | LOG(INFO) << "Begin write thread"; |
1463 | 0 | ASSERT_OK(WaitForBackfillSafeTime(kYBTableName, kIndexName)); |
1464 | |
|
1465 | 0 | LOG(INFO) << "Deleting row"; |
1466 | 0 | ASSERT_OK(conn_->ExecuteFormat("DELETE FROM $0 WHERE i = 1", kTableName)); |
1467 | | |
1468 | | // It should still be in the backfill stage, hopefully before the actual backfill started. |
1469 | 0 | ASSERT_TRUE(ASSERT_RESULT(IsAtTargetIndexStateFlags( |
1470 | 0 | kIndexName, IndexStateFlags{IndexStateFlag::kIndIsLive, IndexStateFlag::kIndIsReady}))); |
1471 | 0 | }); |
1472 | 0 | thread_holder_.JoinAll(); |
1473 | | |
1474 | | // Check. |
1475 | 0 | const Result<PGResultPtr>& result = conn_->FetchFormat( |
1476 | 0 | "SELECT count(*) FROM $0 WHERE j = 'a'", kTableName); |
1477 | 0 | if (result.ok()) { |
1478 | 0 | auto count = ASSERT_RESULT(GetInt64(result.get().get(), 0, 0)); |
1479 | 0 | ASSERT_EQ(count, 0); |
1480 | 0 | } else if (result.status().IsNetworkError()) { |
1481 | 0 | Status s = result.status(); |
1482 | 0 | const std::string msg = s.message().ToBuffer(); |
1483 | 0 | if (msg.find("Given ybctid is not associated with any row in table") == std::string::npos) { |
1484 | 0 | FAIL() << "unexpected status: " << s; |
1485 | 0 | } |
1486 | 0 | FAIL() << "delete to index was not present by the time backfill happened: " << s; |
1487 | 0 | } else { |
1488 | 0 | Status s = result.status(); |
1489 | 0 | FAIL() << "unexpected status: " << s; |
1490 | 0 | } |
1491 | 0 | } |
1492 | | |
1493 | | TEST_F_EX(PgIndexBackfillTest, |
1494 | | YB_DISABLE_TEST_IN_TSAN(IndexScanVisibility), |
1495 | 0 | PgIndexBackfillSlow) { |
1496 | 0 | ExternalTabletServer* diff_ts = cluster_->tablet_server(1); |
1497 | | // Make sure default tserver is 0. At the time of writing, this is set in |
1498 | | // PgWrapperTestBase::SetUp. |
1499 | 0 | ASSERT_NE(pg_ts, diff_ts); |
1500 | |
|
1501 | 0 | LOG(INFO) << "Create connection to run CREATE INDEX"; |
1502 | 0 | PGConn create_index_conn = ASSERT_RESULT(ConnectToDB(kDatabaseName)); |
1503 | 0 | LOG(INFO) << "Create connection to the same tablet server as the one running CREATE INDEX"; |
1504 | 0 | PGConn same_ts_conn = ASSERT_RESULT(ConnectToDB(kDatabaseName)); |
1505 | 0 | LOG(INFO) << "Create connection to a different tablet server from the one running CREATE INDEX"; |
1506 | 0 | PGConn diff_ts_conn = ASSERT_RESULT(PGConn::Connect( |
1507 | 0 | HostPort(diff_ts->bind_host(), diff_ts->pgsql_rpc_port()), |
1508 | 0 | kDatabaseName)); |
1509 | |
|
1510 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE TABLE $0 (i int)", kTableName)); |
1511 | 0 | ASSERT_OK(conn_->ExecuteFormat("INSERT INTO $0 VALUES (1)", kTableName)); |
1512 | |
|
1513 | 0 | thread_holder_.AddThreadFunctor([this, &same_ts_conn, &diff_ts_conn] { |
1514 | 0 | LOG(INFO) << "Begin select thread"; |
1515 | 0 | ASSERT_OK(WaitForBackfillSafeTime(kYBTableName, kIndexName)); |
1516 | |
|
1517 | 0 | LOG(INFO) << "Load DocDB table/index schemas to pggate cache for the other connections"; |
1518 | 0 | ASSERT_RESULT(same_ts_conn.FetchFormat("SELECT * FROM $0 WHERE i = 2", kTableName)); |
1519 | 0 | ASSERT_RESULT(diff_ts_conn.FetchFormat("SELECT * FROM $0 WHERE i = 2", kTableName)); |
1520 | 0 | }); |
1521 | |
|
1522 | 0 | LOG(INFO) << "Create index..."; |
1523 | 0 | ASSERT_OK(create_index_conn.ExecuteFormat("CREATE INDEX $0 ON $1 (i)", kIndexName, kTableName)); |
1524 | 0 | ASSERT_TRUE(thread_holder_.stop_flag()) |
1525 | 0 | << "select thread did not finish by the time CREATE INDEX ended"; |
1526 | 0 | CoarseTimePoint start_time = CoarseMonoClock::Now(); |
1527 | |
|
1528 | 0 | LOG(INFO) << "Check for index scan..."; |
1529 | 0 | const std::string query = Format("SELECT * FROM $0 WHERE i = 2", kTableName); |
1530 | | // The session that ran CREATE INDEX should immediately be ready for index scan. |
1531 | 0 | ASSERT_TRUE(ASSERT_RESULT(create_index_conn.HasIndexScan(query))); |
1532 | | // Eventually, the other sessions should see the index as public. They may take some time because |
1533 | | // they don't know about the latest catalog update until |
1534 | | // 1. master sends catalog version through heartbeat to tserver |
1535 | | // 2. tserver shares catalog version to postgres through shared memory |
1536 | | // Another avenue to learn that the index is public is to send a request to tserver and get a |
1537 | | // schema version mismatch on the indexed table. Since HasIndexScan uses EXPLAIN, it doesn't hit |
1538 | | // tserver, so postgres will be unaware until catalog version is updated in shared memory. Expect |
1539 | | // 0s-1s since default heartbeat period is 1s (see flag heartbeat_interval_ms). |
1540 | 0 | ASSERT_OK(WaitFor( |
1541 | 0 | [&query, &same_ts_conn, &diff_ts_conn]() -> Result<bool> { |
1542 | 0 | bool same_ts_has_index_scan = VERIFY_RESULT(same_ts_conn.HasIndexScan(query)); |
1543 | 0 | bool diff_ts_has_index_scan = VERIFY_RESULT(diff_ts_conn.HasIndexScan(query)); |
1544 | 0 | LOG(INFO) << "same_ts_has_index_scan: " << same_ts_has_index_scan |
1545 | 0 | << ", " |
1546 | 0 | << "diff_ts_has_index_scan: " << diff_ts_has_index_scan; |
1547 | 0 | return same_ts_has_index_scan && diff_ts_has_index_scan; |
1548 | 0 | }, |
1549 | 0 | kIndexStateFlagsUpdateGracePeriod + kIndexStateFlagsUpdateDelay, |
1550 | 0 | "Wait for IndexScan")); |
1551 | 0 | LOG(INFO) << "It took " << yb::ToString(CoarseMonoClock::Now() - start_time) |
1552 | 0 | << " for other sessions to notice that the index became public"; |
1553 | 0 | } |
1554 | | |
1555 | | // Override the index backfill slow test to have smaller WaitUntilIndexPermissionsAtLeast deadline. |
1556 | | class PgIndexBackfillSlowClientDeadline : public PgIndexBackfillSlow { |
1557 | | public: |
1558 | 0 | void UpdateMiniClusterOptions(ExternalMiniClusterOptions* options) override { |
1559 | 0 | PgIndexBackfillSlow::UpdateMiniClusterOptions(options); |
1560 | 0 | options->extra_tserver_flags.push_back(Format( |
1561 | 0 | "--backfill_index_client_rpc_timeout_ms=$0", |
1562 | 0 | (kBackfillDelay / 2).ToMilliseconds())); |
1563 | 0 | } |
1564 | | }; |
1565 | | |
1566 | | // Make sure that the postgres timeout when waiting for backfill to finish causes the index to not |
1567 | | // become public. Simulate the following: |
1568 | | // CREATE INDEX |
1569 | | // - indislive |
1570 | | // - indisready |
1571 | | // - backfill |
1572 | | // - get safe time for read |
1573 | | // - (timeout) |
1574 | | TEST_F_EX(PgIndexBackfillTest, |
1575 | | YB_DISABLE_TEST_IN_TSAN(WaitBackfillTimeout), |
1576 | 0 | PgIndexBackfillSlowClientDeadline) { |
1577 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE TABLE $0 (i int)", kTableName)); |
1578 | 0 | Status status = conn_->ExecuteFormat("CREATE INDEX ON $0 (i)", kTableName); |
1579 | 0 | ASSERT_TRUE(HasClientTimedOut(status)) << status; |
1580 | | |
1581 | | // Make sure that the index is not public. |
1582 | 0 | ASSERT_FALSE(ASSERT_RESULT(conn_->HasIndexScan(Format( |
1583 | 0 | "SELECT * FROM $0 WHERE i = 1", |
1584 | 0 | kTableName)))); |
1585 | 0 | } |
1586 | | |
1587 | | // Make sure that you can still drop an index that failed to fully create. |
1588 | | TEST_F_EX(PgIndexBackfillTest, |
1589 | | YB_DISABLE_TEST_IN_TSAN(DropAfterFail), |
1590 | 0 | PgIndexBackfillSlowClientDeadline) { |
1591 | 0 | auto client = ASSERT_RESULT(cluster_->CreateClient()); |
1592 | 0 | google::protobuf::RepeatedPtrField<master::TabletLocationsPB> tablets; |
1593 | |
|
1594 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE TABLE $0 (i int)", kTableName)); |
1595 | 0 | Status status = conn_->ExecuteFormat("CREATE INDEX $0 ON $1 (i)", kIndexName, kTableName); |
1596 | 0 | ASSERT_TRUE(HasClientTimedOut(status)) << status; |
1597 | | |
1598 | | // Make sure that the index exists in DocDB metadata. |
1599 | 0 | auto tables = ASSERT_RESULT(client->ListTables()); |
1600 | 0 | bool found = false; |
1601 | 0 | for (const auto& table : tables) { |
1602 | 0 | if (table.namespace_name() == kDatabaseName && table.table_name() == kIndexName) { |
1603 | 0 | found = true; |
1604 | 0 | break; |
1605 | 0 | } |
1606 | 0 | } |
1607 | 0 | ASSERT_TRUE(found); |
1608 | |
|
1609 | 0 | ASSERT_OK(conn_->ExecuteFormat("DROP INDEX $0", kIndexName)); |
1610 | | |
1611 | | // Make sure that the index is gone. |
1612 | | // Check postgres metadata. |
1613 | 0 | auto res = ASSERT_RESULT(conn_->FetchFormat( |
1614 | 0 | "SELECT COUNT(*) FROM pg_class WHERE relname = '$0'", kIndexName)); |
1615 | 0 | int64_t value = ASSERT_RESULT(GetInt64(res.get(), 0, 0)); |
1616 | 0 | ASSERT_EQ(value, 0); |
1617 | | // Check DocDB metadata. |
1618 | 0 | tables = ASSERT_RESULT(client->ListTables()); |
1619 | 0 | for (const auto& table : tables) { |
1620 | 0 | ASSERT_FALSE(table.namespace_name() == kDatabaseName && table.table_name() == kIndexName); |
1621 | 0 | } |
1622 | 0 | } |
1623 | | |
1624 | | // Override the index backfill slow test class to have a 30s BackfillIndex client timeout. |
1625 | | class PgIndexBackfillFastClientTimeout : public PgIndexBackfillSlow { |
1626 | | public: |
1627 | 0 | void UpdateMiniClusterOptions(ExternalMiniClusterOptions* options) override { |
1628 | 0 | PgIndexBackfillSlow::UpdateMiniClusterOptions(options); |
1629 | 0 | options->extra_tserver_flags.push_back("--backfill_index_client_rpc_timeout_ms=30000"); |
1630 | 0 | } |
1631 | | }; |
1632 | | |
1633 | | // Make sure that DROP INDEX during backfill is handled well. Simulate the following: |
1634 | | // Session A Session B |
1635 | | // -------------------------- ---------------------- |
1636 | | // CREATE INDEX |
1637 | | // - indislive |
1638 | | // - indisready |
1639 | | // - backfill |
1640 | | // - get safe time for read |
1641 | | // DROP INDEX |
1642 | | TEST_F_EX(PgIndexBackfillTest, |
1643 | | YB_DISABLE_TEST_IN_TSAN(DropWhileBackfilling), |
1644 | 0 | PgIndexBackfillFastClientTimeout) { |
1645 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE TABLE $0 (i int)", kTableName)); |
1646 | | |
1647 | | // conn_ should be used by at most one thread for thread safety. |
1648 | 0 | thread_holder_.AddThreadFunctor([this] { |
1649 | 0 | LOG(INFO) << "Begin create thread"; |
1650 | 0 | PGConn create_conn = ASSERT_RESULT(ConnectToDB(kDatabaseName)); |
1651 | 0 | Status status = create_conn.ExecuteFormat("CREATE INDEX $0 ON $1 (i)", kIndexName, kTableName); |
1652 | | // Expect timeout because |
1653 | | // DROP INDEX is currently not online and removes the index info from the indexed table |
1654 | | // ==> the WaitUntilIndexPermissionsAtLeast will keep failing and retrying GetTableSchema on the |
1655 | | // index. |
1656 | 0 | ASSERT_TRUE(HasClientTimedOut(status)) << status; |
1657 | 0 | }); |
1658 | 0 | thread_holder_.AddThreadFunctor([this] { |
1659 | 0 | LOG(INFO) << "Begin drop thread"; |
1660 | 0 | ASSERT_OK(WaitForBackfillSafeTime(kYBTableName, kIndexName)); |
1661 | |
|
1662 | 0 | LOG(INFO) << "Drop index"; |
1663 | 0 | ASSERT_OK(conn_->ExecuteFormat("DROP INDEX $0", kIndexName)); |
1664 | 0 | }); |
1665 | 0 | thread_holder_.JoinAll(); |
1666 | 0 | } |
1667 | | |
1668 | | // Override the index backfill test class to have a default client admin timeout one second smaller |
1669 | | // than backfill delay. Also, ensure client backfill timeout is high, and set num_tablets to 1 to |
1670 | | // make the test finish more quickly. |
1671 | | class PgIndexBackfillFastDefaultClientTimeout : public PgIndexBackfillTest { |
1672 | | public: |
1673 | 0 | void UpdateMiniClusterOptions(ExternalMiniClusterOptions* options) override { |
1674 | 0 | PgIndexBackfillTest::UpdateMiniClusterOptions(options); |
1675 | 0 | options->extra_tserver_flags.push_back(Format( |
1676 | 0 | "--TEST_slowdown_backfill_by_ms=$0", |
1677 | 0 | kBackfillDelay.ToMilliseconds())); |
1678 | 0 | options->extra_tserver_flags.push_back(Format( |
1679 | 0 | "--yb_client_admin_operation_timeout_sec=$0", (kBackfillDelay - 1s).ToSeconds())); |
1680 | 0 | options->extra_tserver_flags.push_back("--backfill_index_client_rpc_timeout_ms=60000"); // 1m |
1681 | 0 | options->extra_tserver_flags.push_back("--ysql_num_tablets=1"); |
1682 | 0 | } |
1683 | | protected: |
1684 | | const MonoDelta kBackfillDelay = RegularBuildVsSanitizers(7s, 14s); |
1685 | | }; |
1686 | | |
1687 | | // Simply create table and index. The CREATE INDEX should not timeout during backfill because the |
1688 | | // BackfillIndex request from postgres should use the backfill_index_client_rpc_timeout_ms timeout |
1689 | | // (default 60m) rather than the small yb_client_admin_operation_timeout_sec. |
1690 | | TEST_F_EX(PgIndexBackfillTest, |
1691 | | YB_DISABLE_TEST_IN_TSAN(LowerDefaultClientTimeout), |
1692 | 0 | PgIndexBackfillFastDefaultClientTimeout) { |
1693 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE TABLE $0 (i int)", kTableName)); |
1694 | | // This should not time out. |
1695 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE INDEX ON $0 (i)", kTableName)); |
1696 | 0 | } |
1697 | | |
1698 | | // Override the index backfill fast client timeout test class to have more than one master. |
1699 | | class PgIndexBackfillMultiMaster : public PgIndexBackfillFastClientTimeout { |
1700 | | public: |
1701 | 0 | int GetNumMasters() const override { return 3; } |
1702 | | }; |
1703 | | |
1704 | | // Make sure that master leader change during backfill causes the index to not become public and |
1705 | | // doesn't cause any weird hangups or other issues. Simulate the following: |
1706 | | // Session A Session B |
1707 | | // -------------------------- ---------------------- |
1708 | | // CREATE INDEX |
1709 | | // - indislive |
1710 | | // - indisready |
1711 | | // - backfill |
1712 | | // - get safe time for read |
1713 | | // master leader stepdown |
1714 | | // TODO(jason): update this test when handling master leader changes during backfill (issue #6218). |
1715 | | TEST_F_EX(PgIndexBackfillTest, |
1716 | | YB_DISABLE_TEST_IN_TSAN(MasterLeaderStepdown), |
1717 | 0 | PgIndexBackfillMultiMaster) { |
1718 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE TABLE $0 (i int)", kTableName)); |
1719 | | |
1720 | | // conn_ should be used by at most one thread for thread safety. |
1721 | 0 | thread_holder_.AddThreadFunctor([this] { |
1722 | 0 | LOG(INFO) << "Begin create thread"; |
1723 | 0 | PGConn create_conn = ASSERT_RESULT(ConnectToDB(kDatabaseName)); |
1724 | | // The CREATE INDEX should get master leader change during backfill so that its |
1725 | | // WaitUntilIndexPermissionsAtLeast call starts querying the new leader. Since the new leader |
1726 | | // will be inactive at the WRITE_AND_DELETE docdb permission, it will wait until the deadline, |
1727 | | // which is set to 30s. |
1728 | 0 | Status status = create_conn.ExecuteFormat("CREATE INDEX $0 ON $1 (i)", kIndexName, kTableName); |
1729 | 0 | ASSERT_TRUE(HasClientTimedOut(status)) << status; |
1730 | 0 | }); |
1731 | 0 | thread_holder_.AddThreadFunctor([this] { |
1732 | 0 | LOG(INFO) << "Begin master leader stepdown thread"; |
1733 | 0 | ASSERT_OK(WaitForBackfillSafeTime(kYBTableName, kIndexName)); |
1734 | |
|
1735 | 0 | LOG(INFO) << "Doing master leader stepdown"; |
1736 | 0 | tserver::TabletServerErrorPB::Code error_code; |
1737 | 0 | ASSERT_OK(cluster_->StepDownMasterLeader(&error_code)); |
1738 | | |
1739 | | // It should still be in the backfill stage, hopefully before the actual backfill started. |
1740 | 0 | ASSERT_TRUE(ASSERT_RESULT(IsAtTargetIndexStateFlags( |
1741 | 0 | kIndexName, IndexStateFlags{IndexStateFlag::kIndIsLive, IndexStateFlag::kIndIsReady}))); |
1742 | 0 | }); |
1743 | 0 | thread_holder_.JoinAll(); |
1744 | 0 | } |
1745 | | |
1746 | | // Override the index backfill test class to use colocated tables. |
1747 | | class PgIndexBackfillColocated : public PgIndexBackfillTest { |
1748 | | public: |
1749 | 0 | void SetUp() override { |
1750 | 0 | LibPqTestBase::SetUp(); |
1751 | |
|
1752 | 0 | PGConn conn_init = ASSERT_RESULT(Connect()); |
1753 | 0 | ASSERT_OK(conn_init.ExecuteFormat("CREATE DATABASE $0 WITH colocated = true", kColoDbName)); |
1754 | |
|
1755 | 0 | conn_ = std::make_unique<PGConn>(ASSERT_RESULT(ConnectToDB(kColoDbName))); |
1756 | 0 | } |
1757 | | }; |
1758 | | |
1759 | | // Make sure that backfill works when colocation is on. |
1760 | | TEST_F_EX(PgIndexBackfillTest, |
1761 | | YB_DISABLE_TEST_IN_TSAN(ColocatedSimple), |
1762 | 0 | PgIndexBackfillColocated) { |
1763 | 0 | TestSimpleBackfill(); |
1764 | 0 | } |
1765 | | |
1766 | | // Make sure that backfill works when there are multiple colocated tables. |
1767 | | TEST_F_EX(PgIndexBackfillTest, |
1768 | | YB_DISABLE_TEST_IN_TSAN(ColocatedMultipleTables), |
1769 | 0 | PgIndexBackfillColocated) { |
1770 | | // Create two tables with the index on the second table. |
1771 | 0 | const std::string kOtherTable = "yyy"; |
1772 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE TABLE $0 (i int)", kOtherTable)); |
1773 | 0 | ASSERT_OK(conn_->ExecuteFormat("INSERT INTO $0 VALUES (100)", kOtherTable)); |
1774 | |
|
1775 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE TABLE $0 (i int)", kTableName)); |
1776 | 0 | ASSERT_OK(conn_->ExecuteFormat("INSERT INTO $0 VALUES (200)", kTableName)); |
1777 | 0 | ASSERT_OK(conn_->ExecuteFormat("INSERT INTO $0 VALUES (300)", kTableName)); |
1778 | 0 | ASSERT_OK(conn_->ExecuteFormat("CREATE INDEX ON $0 (i ASC)", kTableName)); |
1779 | | |
1780 | | // Index scan to verify contents of index table. |
1781 | 0 | const std::string query = Format("SELECT COUNT(*) FROM $0 WHERE i > 0", kTableName); |
1782 | 0 | ASSERT_TRUE(ASSERT_RESULT(conn_->HasIndexScan(query))); |
1783 | 0 | PGResultPtr res = ASSERT_RESULT(conn_->Fetch(query)); |
1784 | 0 | auto count = ASSERT_RESULT(GetInt64(res.get(), 0, 0)); |
1785 | 0 | ASSERT_EQ(count, 2); |
1786 | 0 | } |
1787 | | |
1788 | | // Test that retain_delete_markers is properly set after index backfill for a colocated table. |
1789 | | TEST_F_EX(PgIndexBackfillTest, |
1790 | | YB_DISABLE_TEST_IN_TSAN(ColocatedRetainDeleteMarkers), |
1791 | 0 | PgIndexBackfillColocated) { |
1792 | 0 | TestRetainDeleteMarkers(kColoDbName); |
1793 | 0 | } |
1794 | | |
1795 | | } // namespace pgwrapper |
1796 | | } // namespace yb |