----------------------------------------------------------------------- WORKLOG TASK -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- TASK...........: mysqlbinlog: remove undesired effects of format description binlog statement CREATION DATE..: Mon, 17 Aug 2009, 14:12 SUPERVISOR.....: Monty IMPLEMENTOR....: COPIES TO......: CATEGORY.......: Server-RawIdeaBin TASK ID........: 50 (http://askmonty.org/worklog/?tid=50) VERSION........: Server-9.x STATUS.........: Un-Assigned PRIORITY.......: 60 WORKED HOURS...: 14 ESTIMATE.......: 6 (hours remain) ORIG. ESTIMATE.: 10 PROGRESS NOTES: -=-=(Knielsen - Mon, 28 Sep 2009, 16:00)=-=- Updated patch against MySQL 5.1.39 Fixes a couple of issues with the patch. Discussions with MySQL developer. Worked 4 hours and estimate 6 hours remain (original estimate unchanged). -=-=(Knielsen - Wed, 09 Sep 2009, 18:31)=-=- Discussed with MySQL developers on Bug#46640. Committed a suggested fix: https://lists.launchpad.net/maria-developers/msg00926.html Worked 8 hours and estimate 10 hours remain (original estimate increased by 18 hours). -=-=(Guest - Wed, 09 Sep 2009, 11:50)=-=- High Level Description modified. --- /tmp/wklog.50.old.11584 2009-09-09 11:50:12.000000000 +0300 +++ /tmp/wklog.50.new.11584 2009-09-09 11:50:12.000000000 +0300 @@ -5,7 +5,11 @@ This will cause an error when one tires to apply mysqlbinlog output to a 5.0.x server. -2. When one applies "format description binlog statement" at the slave, it -will change the slave's server_id when applied. +2. When one executes a BINLOG statement (containing "format description binlog +statement" or other events), this will cause subsequent statements run in +the same session to be binlogged with the server id of the last event +executed, not with the real server id. This can cause problems like infinite +recursive application in a circular replication topology. This WL is to fix these issues. + -=-=(Knielsen - Tue, 08 Sep 2009, 15:07)=-=- Low Level Design modified. --- /tmp/wklog.50.old.19208 2009-09-08 15:07:09.000000000 +0300 +++ /tmp/wklog.50.new.19208 2009-09-08 15:07:09.000000000 +0300 @@ -1 +1,34 @@ +I think the fix for point 2 is to replace the call to +apply_event_and_update_pos() in mysql_client_binlog_statement() with a direct +call to ev->apply_event(). However, I need to check two things to be sure this +is correct: + +1. The existing code does + + if (!ev->when) ev->when= my_time(0) + +Need to understand if this is needed/correct or not. + +2. The existing code does ev->update_pos(). I think this is redundant for +BINLOG statement (as it uses a fake rli structure), but I need to check to +make sure. + +Once this is done, point 1 may no longer be needed. The user can use +--base64-output=never when applying the output to a 5.0 server, and omit that +option when applying the output to a 5.1 server. There should be no need to +omit the format description event in the output when other BINLOG statements +for row events are present, as it no longer changes the server id. In fact the +format description event is required to be able to execute other BINLOG +statements, as its purpose is to define the binary format of the events +contained in these statements. + +Alternatively, we could implement that if the format description event of the +source binlog has server version < 5.1, and --base64-output=auto (default), +then the format description event is omitted (and should any BINLOG statement +be needed, unlikely as it is from a 5.0 server, we will need to throw an +error). + +The binlog format version for 5.0 and 5.1 is actually the same, hence the need +to look at server version to guess if the format description event can be +omitted. -=-=(Knielsen - Tue, 08 Sep 2009, 14:29)=-=- High-Level Specification modified. --- /tmp/wklog.50.old.17531 2009-09-08 14:29:30.000000000 +0300 +++ /tmp/wklog.50.new.17531 2009-09-08 14:29:30.000000000 +0300 @@ -12,3 +12,37 @@ One question that one needs to sort out before disabling server_id change is why it was put there in the first place? Should it be always removed? +First, the problem is not the format description log event. In fact all log +events have their own value of server_id, and every event in the BINLOG +statement is executed with its own id as server_id. + +It seems it was introduced deliberately in the patch for Bug#32407. This patch +makes sql thread and BINLOG statement share code path for executing the binary +event, even though the bug is really about something different (outputting +format description event to allow proper execution of other BINLOG +statements). + +Nevertheless, I think using the server_id in the events of a BINLOG statement +is wrong: + +1. It is different behaviour from executing normal statement-based events not +written using BINLOG. + +2. It causes the possibility for infinite cycle of replication in a cyclic +replication setup, if the server ids in BINLOG are different from all other +server ids in the cycle. + +3. The functionality of applying events with original server id from the event +is still available by pointing the slave thread to the binlog. The +functionality of applying row-based binlog events with server id of the server +executing the BINLOG statements is not otherwise available. + +In fact most of the code from the slave thread that is now also run when +executing BINLOG statements is redundant, or does the wrong thing (like +updating binlog position). + +Maybe there is a confusion on the purpose of the BINLOG statement. Is it to be +able to apply row-based events manually to a server (mysqlbinlog|mysql)? Or is +it to be able to simulate the effect of the slave sql thread? I think the +former is the most important one. + -=-=(Knielsen - Mon, 07 Sep 2009, 10:06)=-=- Research worklog, bugs, and existing server code. Worked 2 hours and estimate 0 hours remain (original estimate increased by 2 hours). -=-=(Psergey - Mon, 17 Aug 2009, 14:13)=-=- Dependency created: 39 now depends on 50 -=-=(Psergey - Mon, 17 Aug 2009, 14:13)=-=- High-Level Specification modified. --- /tmp/wklog.50.old.11389 2009-08-17 14:13:05.000000000 +0300 +++ /tmp/wklog.50.new.11389 2009-08-17 14:13:05.000000000 +0300 @@ -1 +1,14 @@ +First item +---------- +AFAIU what needs to be done is: +1. record a source server version (it is in the first binlog event). +2. don't emit a BINLOG statement if the recorded version number is 5.0.x or + below. + +and we'll get a 5.0-applicable binlog. + +Second item +----------- +One question that one needs to sort out before disabling server_id change is +why it was put there in the first place? Should it be always removed? DESCRIPTION: According to complaints in BUG#46640: 1. mysqlbinlog will emit a 5.0-incompatible "format description binlog statement" even when reading a binary log that was produced by 5.0. This will cause an error when one tires to apply mysqlbinlog output to a 5.0.x server. 2. When one executes a BINLOG statement (containing "format description binlog statement" or other events), this will cause subsequent statements run in the same session to be binlogged with the server id of the last event executed, not with the real server id. This can cause problems like infinite recursive application in a circular replication topology. This WL is to fix these issues. HIGH-LEVEL SPECIFICATION: First item ---------- AFAIU what needs to be done is: 1. record a source server version (it is in the first binlog event). 2. don't emit a BINLOG statement if the recorded version number is 5.0.x or below. and we'll get a 5.0-applicable binlog. Second item ----------- One question that one needs to sort out before disabling server_id change is why it was put there in the first place? Should it be always removed? First, the problem is not the format description log event. In fact all log events have their own value of server_id, and every event in the BINLOG statement is executed with its own id as server_id. It seems it was introduced deliberately in the patch for Bug#32407. This patch makes sql thread and BINLOG statement share code path for executing the binary event, even though the bug is really about something different (outputting format description event to allow proper execution of other BINLOG statements). Nevertheless, I think using the server_id in the events of a BINLOG statement is wrong: 1. It is different behaviour from executing normal statement-based events not written using BINLOG. 2. It causes the possibility for infinite cycle of replication in a cyclic replication setup, if the server ids in BINLOG are different from all other server ids in the cycle. 3. The functionality of applying events with original server id from the event is still available by pointing the slave thread to the binlog. The functionality of applying row-based binlog events with server id of the server executing the BINLOG statements is not otherwise available. In fact most of the code from the slave thread that is now also run when executing BINLOG statements is redundant, or does the wrong thing (like updating binlog position). Maybe there is a confusion on the purpose of the BINLOG statement. Is it to be able to apply row-based events manually to a server (mysqlbinlog|mysql)? Or is it to be able to simulate the effect of the slave sql thread? I think the former is the most important one. LOW-LEVEL DESIGN: I think the fix for point 2 is to replace the call to apply_event_and_update_pos() in mysql_client_binlog_statement() with a direct call to ev->apply_event(). However, I need to check two things to be sure this is correct: 1. The existing code does if (!ev->when) ev->when= my_time(0) Need to understand if this is needed/correct or not. 2. The existing code does ev->update_pos(). I think this is redundant for BINLOG statement (as it uses a fake rli structure), but I need to check to make sure. Once this is done, point 1 may no longer be needed. The user can use --base64-output=never when applying the output to a 5.0 server, and omit that option when applying the output to a 5.1 server. There should be no need to omit the format description event in the output when other BINLOG statements for row events are present, as it no longer changes the server id. In fact the format description event is required to be able to execute other BINLOG statements, as its purpose is to define the binary format of the events contained in these statements. Alternatively, we could implement that if the format description event of the source binlog has server version < 5.1, and --base64-output=auto (default), then the format description event is omitted (and should any BINLOG statement be needed, unlikely as it is from a 5.0 server, we will need to throw an error). The binlog format version for 5.0 and 5.1 is actually the same, hence the need to look at server version to guess if the format description event can be omitted. ESTIMATED WORK TIME ESTIMATED COMPLETION DATE ----------------------------------------------------------------------- WorkLog (v3.5.9)