Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
Run catalog reindexing test from 3dbb317d32 serially, to avoid deadlocks.
authorAndres Freund <andres@anarazel.de>
Wed, 1 May 2019 00:45:32 +0000 (17:45 -0700)
committerAndres Freund <andres@anarazel.de>
Wed, 1 May 2019 00:45:32 +0000 (17:45 -0700)
The tests turn out to cause deadlocks in some circumstances. Fairly
reproducibly so with -DRELCACHE_FORCE_RELEASE
-DCATCACHE_FORCE_RELEASE.  Some of the deadlocks may be hard to fix
without disproportionate measures, but others probably should be fixed
- but not in 12.

We discussed removing the new tests until we can fix the issues
underlying the deadlocks, but results from buildfarm animal
markhor (which runs with CLOBBER_CACHE_ALWAYS) indicates that there
might be a more severe, as of yet undiagnosed, issue (including on
stable branches) with reindexing catalogs. The failure is:
ERROR: could not read block 0 in file "base/16384/28025": read only 0 of 8192 bytes
Therefore it seems advisable to keep the tests.

It's not certain that running the tests in isolation removes the risk
of deadlocks. It's possible that additional locks are needed to
protect against a concurrent auto-analyze or such.

Per discussion with Tom Lane.

Discussion: https://postgr.es/m/28926.1556664156@sss.pgh.pa.us
Backpatch: 9.4-, like 3dbb317d3

src/test/regress/expected/create_index.out
src/test/regress/expected/reindex_catalog.out [new file with mode: 0644]
src/test/regress/parallel_schedule
src/test/regress/serial_schedule
src/test/regress/sql/create_index.sql
src/test/regress/sql/reindex_catalog.sql [new file with mode: 0644]

index fa6d86bbd4ad0a419bb1721a156348516023a0e3..be25101db24cbbc673aec6230e436dcf44eca5d8 100644 (file)
@@ -3071,24 +3071,6 @@ REINDEX (VERBOSE) TABLE reindex_verbose;
 INFO:  index "reindex_verbose_pkey" was reindexed
 DROP TABLE reindex_verbose;
 --
--- check that system tables can be reindexed
---
--- whole tables
-REINDEX TABLE pg_class; -- mapped, non-shared, critical
-REINDEX TABLE pg_index; -- non-mapped, non-shared, critical
-REINDEX TABLE pg_operator; -- non-mapped, non-shared, critical
-REINDEX TABLE pg_database; -- mapped, shared, critical
-REINDEX TABLE pg_shdescription; -- mapped, shared non-critical
--- Check that individual system indexes can be reindexed. That's a bit
--- different from the entire-table case because reindex_relation
--- treats e.g. pg_class special.
-REINDEX INDEX pg_class_oid_index; -- mapped, non-shared, critical
-REINDEX INDEX pg_class_relname_nsp_index; -- mapped, non-shared, non-critical
-REINDEX INDEX pg_index_indexrelid_index; -- non-mapped, non-shared, critical
-REINDEX INDEX pg_index_indrelid_index; -- non-mapped, non-shared, non-critical
-REINDEX INDEX pg_database_oid_index; -- mapped, shared, critical
-REINDEX INDEX pg_shdescription_o_c_index; -- mapped, shared, non-critical
---
 -- REINDEX SCHEMA
 --
 REINDEX SCHEMA schema_to_reindex; -- failure, schema does not exist
diff --git a/src/test/regress/expected/reindex_catalog.out b/src/test/regress/expected/reindex_catalog.out
new file mode 100644 (file)
index 0000000..142616f
--- /dev/null
@@ -0,0 +1,33 @@
+--
+-- Check that system tables can be reindexed.
+--
+-- Note that this test currently has to run without parallel tests
+-- being scheduled, as currently reindex catalog tables can cause
+-- deadlocks:
+--
+-- * The lock upgrade between the ShareLock acquired for the reindex
+--   and RowExclusiveLock needed for pg_class/pg_index locks can
+--   trigger deadlocks.
+--
+-- * The uniqueness checks performed when reindexing a unique/primary
+--   key index possibly need to wait for the transaction of a
+--   about-to-deleted row in pg_class to commit. That can cause
+--   deadlocks because, in contrast to user tables, locks on catalog
+--   tables are routinely released before commit - therefore the lock
+--   held for reindexing doesn't guarantee that no running transaction
+--   performed modifications in the table underlying the index.
+-- Check reindexing of whole tables
+REINDEX TABLE pg_class; -- mapped, non-shared, critical
+REINDEX TABLE pg_index; -- non-mapped, non-shared, critical
+REINDEX TABLE pg_operator; -- non-mapped, non-shared, critical
+REINDEX TABLE pg_database; -- mapped, shared, critical
+REINDEX TABLE pg_shdescription; -- mapped, shared non-critical
+-- Check that individual system indexes can be reindexed. That's a bit
+-- different from the entire-table case because reindex_relation
+-- treats e.g. pg_class special.
+REINDEX INDEX pg_class_oid_index; -- mapped, non-shared, critical
+REINDEX INDEX pg_class_relname_nsp_index; -- mapped, non-shared, non-critical
+REINDEX INDEX pg_index_indexrelid_index; -- non-mapped, non-shared, critical
+REINDEX INDEX pg_index_indrelid_index; -- non-mapped, non-shared, non-critical
+REINDEX INDEX pg_database_oid_index; -- mapped, shared, critical
+REINDEX INDEX pg_shdescription_o_c_index; -- mapped, shared, non-critical
index 8112626f3bd594623954cfc5ee518f5d516f605c..cdbb08c7b3e9e66cf508849dff79a18d3af11902 100644 (file)
@@ -57,6 +57,11 @@ test: create_misc create_operator create_procedure
 # These depend on the above two
 test: create_index create_view index_including
 
+# ----------
+# Has to run in isolation, due to deadlock risk
+# ----------
+test: reindex_catalog
+
 # ----------
 # Another group of parallel tests
 # ----------
index b2a8f3705630d8ce459387690ee0b93d87469253..7de53c617f861ac9980ed007185acba859edd006 100644 (file)
@@ -67,6 +67,7 @@ test: create_procedure
 test: create_index
 test: index_including
 test: create_view
+test: reindex_catalog
 test: create_aggregate
 test: create_function_3
 test: create_cast
index 121c78c4bfdadd8cde741eb08665caea3925441a..f9e7118f0d3d864325de765f83e6015f90bbd778 100644 (file)
@@ -1080,27 +1080,6 @@ CREATE TABLE reindex_verbose(id integer primary key);
 REINDEX (VERBOSE) TABLE reindex_verbose;
 DROP TABLE reindex_verbose;
 
---
--- check that system tables can be reindexed
---
-
--- whole tables
-REINDEX TABLE pg_class; -- mapped, non-shared, critical
-REINDEX TABLE pg_index; -- non-mapped, non-shared, critical
-REINDEX TABLE pg_operator; -- non-mapped, non-shared, critical
-REINDEX TABLE pg_database; -- mapped, shared, critical
-REINDEX TABLE pg_shdescription; -- mapped, shared non-critical
-
--- Check that individual system indexes can be reindexed. That's a bit
--- different from the entire-table case because reindex_relation
--- treats e.g. pg_class special.
-REINDEX INDEX pg_class_oid_index; -- mapped, non-shared, critical
-REINDEX INDEX pg_class_relname_nsp_index; -- mapped, non-shared, non-critical
-REINDEX INDEX pg_index_indexrelid_index; -- non-mapped, non-shared, critical
-REINDEX INDEX pg_index_indrelid_index; -- non-mapped, non-shared, non-critical
-REINDEX INDEX pg_database_oid_index; -- mapped, shared, critical
-REINDEX INDEX pg_shdescription_o_c_index; -- mapped, shared, non-critical
-
 --
 -- REINDEX SCHEMA
 --
diff --git a/src/test/regress/sql/reindex_catalog.sql b/src/test/regress/sql/reindex_catalog.sql
new file mode 100644 (file)
index 0000000..2180ee5
--- /dev/null
@@ -0,0 +1,36 @@
+--
+-- Check that system tables can be reindexed.
+--
+-- Note that this test currently has to run without parallel tests
+-- being scheduled, as currently reindex catalog tables can cause
+-- deadlocks:
+--
+-- * The lock upgrade between the ShareLock acquired for the reindex
+--   and RowExclusiveLock needed for pg_class/pg_index locks can
+--   trigger deadlocks.
+--
+-- * The uniqueness checks performed when reindexing a unique/primary
+--   key index possibly need to wait for the transaction of a
+--   about-to-deleted row in pg_class to commit. That can cause
+--   deadlocks because, in contrast to user tables, locks on catalog
+--   tables are routinely released before commit - therefore the lock
+--   held for reindexing doesn't guarantee that no running transaction
+--   performed modifications in the table underlying the index.
+
+
+-- Check reindexing of whole tables
+REINDEX TABLE pg_class; -- mapped, non-shared, critical
+REINDEX TABLE pg_index; -- non-mapped, non-shared, critical
+REINDEX TABLE pg_operator; -- non-mapped, non-shared, critical
+REINDEX TABLE pg_database; -- mapped, shared, critical
+REINDEX TABLE pg_shdescription; -- mapped, shared non-critical
+
+-- Check that individual system indexes can be reindexed. That's a bit
+-- different from the entire-table case because reindex_relation
+-- treats e.g. pg_class special.
+REINDEX INDEX pg_class_oid_index; -- mapped, non-shared, critical
+REINDEX INDEX pg_class_relname_nsp_index; -- mapped, non-shared, non-critical
+REINDEX INDEX pg_index_indexrelid_index; -- non-mapped, non-shared, critical
+REINDEX INDEX pg_index_indrelid_index; -- non-mapped, non-shared, non-critical
+REINDEX INDEX pg_database_oid_index; -- mapped, shared, critical
+REINDEX INDEX pg_shdescription_o_c_index; -- mapped, shared, non-critical