mdev-31272 new fix variant #332
Conversation
sjaakola
force-pushed
the
10.6-MDEV-31272-new
branch
from
3d04fe2 to
9b32f40
Compare
If a statement execution has successfully modified some rows, then the write set for the transaction, will be appended by the key values related to the modified rows. If the staement execution eventually fails, e.g. for duplicate key error, then statement rollback will be carried out, to undo changes in storage engine and truncate binlog cache to wipe out replication events related to the statement execution. If the transaction has no other successful write statements then there is nothing to commit, at commit time, and no binlog events either. But as there were earlier key appends, the transaction remains registred for wsrep replication, and at commit time replicator extracts binlog events from cache and populates the final write set for the replication. In the buggy version, it was observed that there are no binlog events, but the write set was populated, nevertheless, and the empty write set was replicated. Note that this consumes one GTID, for a void operation. The fix for this issue, cancels the replication of such empty write set anf gives a warning for the client. This fix is in: Wsrep_client_service::prepare_data_for_replication() It also turned out that the binlog cache truncate for statement rollback was not complete if server was not configured to use binloggging. In buggy version, only statmenet cache was truncated, but events remained in the transaction cache. This has been fixed in wsrep_thd_binlog_stmt_rollback()
binlog_commit maintains also binlog cache, which is needed also when binlogging is not used, e.g. to support steement rollback
…have empty last WS carrying only commit flag
methods for managing binlog cache when actual binlog is not used
…ines mtr test galera.galera_ctas should pass with this patch
where IO cache is looked and if no events in cache, then wsrep commit hooks will be skipped. With this, wsrep replication will not be tried, and first version of the fix becomes obsolete.
sjaakola
force-pushed
the
10.6-MDEV-31272-new
branch
from
9b32f40 to
5e29b75
Compare
| @@ -138,6 +136,10 @@ int Wsrep_client_service::prepare_data_for_replication() | |||
| { | |||
| WSREP_DEBUG("empty rbr buffer, query: %s", wsrep_thd_query(m_thd)); | |||
| } | |||
|
|
|||
| /* SR may have empty last WS to just carry the comit flag */ | |||
| @@ -375,7 +398,7 @@ static inline int wsrep_after_commit(THD* thd, bool all) | |||
| wsrep_is_active(thd), | |||
| (long long)wsrep_thd_trx_seqno(thd), | |||
| wsrep_has_changes(thd)); | |||
| DBUG_ASSERT(wsrep_run_commit_hook(thd, all)); | |||
| //DBUG_ASSERT(wsrep_run_commit_hook(thd, all)); | |||
There was a problem hiding this comment.
Let's not have assertions that are commented out. Either remove this line altogether, or put it back and fix if necessary.
| if (thd->wsrep_cs().mode() == Wsrep_client_state::m_local && | ||
| !thd->wsrep_trx().is_streaming()) | ||
| { | ||
| IO_CACHE* trx_cache= wsrep_get_cache(thd, true); |
There was a problem hiding this comment.
Perhaps it would be good to put this block into a function in wsrep_binlog.cc or log.cc
| @@ -348,7 +348,7 @@ class binlog_cache_data | |||
|
|
|||
| my_off_t get_prev_position() | |||
| { | |||
| return(before_stmt_pos); | |||
| up the cache, so this must be done explicitly when the transaction | ||
| terminates. | ||
| */ | ||
| if (WSREP_EMULATE_BINLOG_NNULL(this)) |
There was a problem hiding this comment.
If you drop this optimization, then we should probably remove the patch at the beginning of ha_savepoint().
|
|
||
|
|
||
| # | ||
| # Case 5: testing statement rollback |
There was a problem hiding this comment.
Case 5: It is not so clear what case 5 is trying to test. If I understand correctly, you want to make sure that the COMMIT from node_1 does not go through replication / certification because it is an empty transaction. If so simply check that no gtid was consumed, or alternatively that seqno has not advanced. This case would not need debug sync at all and make it trivial to understand.
| DROP TABLE t1; | ||
|
|
||
| # | ||
| # Case 7: |
There was a problem hiding this comment.
Case 7 looks like case 2. If so, drop it.
|
|
||
|
|
||
| # | ||
| # Case 6: testing statement rollback with BF abort |
There was a problem hiding this comment.
Case 6 can also simplified and make use of wait conditions instead of debug sync points (see galera_bf_abort.test). If we make this test debug sync free it has the advantage of being run in debug and release builds.
Here, empty transaction is checked in wsrep_run_commit_hook