Обсуждение: BRIN summarization vs. WAL logging
Hi, In a thread about sequences and sync replication [1], I've explained that the issue we're observing is due to not waiting for WAL at commit if the transaction only did nextval(). In which case we don't flush WAL in RecordTransactionCommit, we don't wait for sync replica, etc. The WAL may get lost in case of crash, etc. As I explained in the other thread, there are various other cases where a transaction generates WAL but does not have XID, which is sufficient for not flushing/waiting at transaction commit. Some of those cases are probably fine (e.g. there are comments explaining why this is fine for PRUNE record). But other cases (discovered by running regression tests with extra logging) looked a bit suspicious - particularly those that write multiple WAL messages, because what if we lose just some of those? So I looked at two cases related to BRIN, mostly because those were fairly simple, and I think at least the brin_summarize_range() is somewhat broken. 1) brin_desummarize_range() This is pretty simple, because this function generates a single WAL record, without waiting for it to be flushed: DESUMMARIZE pagesPerRange 1, heapBlk 0, page offset 9, blkref #0: ... But if the cluster/VM/... crashes right after you ran the function (and it completed just fine, possibly even in an explicit transaciton), that change will get lost. Not really a serious data corruption/loss, and you can simply run it again, but IMHO rather surprising. Of course, most people are unlikely to run brin_desummarize_range() very often, so maybe it's acceptable? But of course - if we expect this to be very rare operation, why skip the WAL at all? 2) brin_summarize_range() Now, the issue I think is more serious, more likely to happen, and harder to fix. When summarizing a range, we write two WAL records: INSERT heapBlk 2 pagesPerRange 2 offnum 2, blkref #0: rel 1663/63 ... SAMEPAGE_UPDATE offnum 2, blkref #0: rel 1663/63341/73957 blk 2 So, what happens if we lost the second WAL record, e.g. due to a crash? To experiment with this, I wrote a trivial patch (attached) that allows crashing on WAL message of certain type by simply setting a GUC. Now, consider this example: create table t (a int); insert into t select i from generate_series(1,5000) s(i); create index on t using brin (a); select brin_desummarize_range('t_a_idx', 1); set crash_on_wal_message = 'SAMEPAGE_UPDATE'; select brin_summarize_range('t_a_idx', 5); PANIC: crashing before 'SAMEPAGE_UPDATE' WAL message server closed the connection unexpectedly ... After recovery, this is what we have: select * from brin_page_items(get_Raw_page('t_a_idx', 2), 't_a_idx'); ... | allnulls | hasnulls | placeholder | value ... -+----------+----------+-------------+------- ... | t | f | t | (1 row) So the BRIN tuple is still marked as placeholder, which is a problem because that means we'll always consider it as matching, making the bitmap index scan less efficient. And we'll *never* fix this, because just summarizing the range does nothing: select brin_summarize_range('t_a_idx', 5); brin_summarize_range ---------------------- 0 (1 row) So it's still marked as placeholder, and to fix it you have to explicitly desummarize the range first. The reason for this seems obvious - only the process that created the placeholder tuple is expected to mark it as "placeholder=false", but this is described as two WAL records. And if we lose the update, the tuple will stay marked as a placeholder forever. Of course, this requires a crash while something is summarizing ranges. But consider the summarization is often done by autovacuum, so it's not just about hitting this from manually-executed brin_summarize_range. I'm not quite sure what to do about this. Just doing XLogFlush() does not really fix this - it makes it less likely, but the root cause is the change is described by multiple WAL messages that are not linked together in any way. We may lost the last message without noticing that, and the flush does not fix that. I didn't look at the other cases mentioned in [1], but I would't be surprised if some had a similar issue (e.g. the GIN pending list cleanup seems like another candidate). regards [1] https://www.postgresql.org/message-id/0f827a71-a01b-bcf9-fe77-3047a9d4a93c%40enterprisedb.com -- Tomas Vondra EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
Вложения
On Tue, Jan 25, 2022 at 10:12 PM Tomas Vondra <tomas.vondra@enterprisedb.com> wrote: > 1) brin_desummarize_range() > > But if the cluster/VM/... crashes right after you ran the function (and > it completed just fine, possibly even in an explicit transaciton), that > change will get lost. Not really a serious data corruption/loss, and you > can simply run it again, but IMHO rather surprising. I don't have a big problem with inserting an XLogFlush() here, but I also don't think there's a hard and fast rule that maintenance operations have to have full transactional behavior. Because that's violated in various different ways by various different DDL commands. CREATE INDEX CONCURRENTLY can leave detritus around if it fails. At least one variant of ALTER TABLE does a non-MVCC-aware rewrite of the table contents. COPY FREEZE violates MVCC. Concurrent partition attach and detach aren't fully serializable with concurrent transactions. Years ago TRUNCATE didn't have MVCC semantics. We often make small compromises in these areas for implementation simplicity or other benefits. > 2) brin_summarize_range() > > Now, the issue I think is more serious, more likely to happen, and > harder to fix. When summarizing a range, we write two WAL records: > > INSERT heapBlk 2 pagesPerRange 2 offnum 2, blkref #0: rel 1663/63 ... > SAMEPAGE_UPDATE offnum 2, blkref #0: rel 1663/63341/73957 blk 2 > > So, what happens if we lost the second WAL record, e.g. due to a crash? Ouch. As you say, XLogFlush() won't fix that. The WAL logging scheme needs to be redesigned. -- Robert Haas EDB: http://www.enterprisedb.com
On 2022-Jan-26, Robert Haas wrote: > On Tue, Jan 25, 2022 at 10:12 PM Tomas Vondra > <tomas.vondra@enterprisedb.com> wrote: > > 2) brin_summarize_range() > > > > Now, the issue I think is more serious, more likely to happen, and > > harder to fix. When summarizing a range, we write two WAL records: > > > > INSERT heapBlk 2 pagesPerRange 2 offnum 2, blkref #0: rel 1663/63 ... > > SAMEPAGE_UPDATE offnum 2, blkref #0: rel 1663/63341/73957 blk 2 > > > > So, what happens if we lost the second WAL record, e.g. due to a crash? > > Ouch. As you say, XLogFlush() won't fix that. The WAL logging scheme > needs to be redesigned. I'm not sure what a good fix is. I was thinking that maybe if a placeholder tuple is found during index scan, and the corresponding process is no longer running, then the index scanner would remove the placeholder tuple, making the range unsummarized again. However, how would we know that the process is gone? Another idea is to use WAL's rm_cleanup: during replay, remember that a placeholder tuple was seen, then remove the info if we see an update from the originating process that replaces the placeholder tuple with a real one; at cleanup time, if the list of remembered placeholder tuples is nonempty, remove them. (I vaguely recall we used the WAL rm_cleanup mechanism for something like this, but we no longer do AFAICS.) ... Oh, but if there is a promotion involved, we may end up with a placeholder insertion before the promotion and the update afterwards. That would probably not be handled well. One thing not completely clear to me is whether this only affects placeholder tuples. Is it possible to have this problem with regular BRIN tuples? I think it isn't. -- Álvaro Herrera Valdivia, Chile — https://www.EnterpriseDB.com/
On 1/26/22 19:14, Alvaro Herrera wrote: > On 2022-Jan-26, Robert Haas wrote: > >> On Tue, Jan 25, 2022 at 10:12 PM Tomas Vondra >> <tomas.vondra@enterprisedb.com> wrote: > >>> 2) brin_summarize_range() >>> >>> Now, the issue I think is more serious, more likely to happen, and >>> harder to fix. When summarizing a range, we write two WAL records: >>> >>> INSERT heapBlk 2 pagesPerRange 2 offnum 2, blkref #0: rel 1663/63 ... >>> SAMEPAGE_UPDATE offnum 2, blkref #0: rel 1663/63341/73957 blk 2 >>> >>> So, what happens if we lost the second WAL record, e.g. due to a crash? >> >> Ouch. As you say, XLogFlush() won't fix that. The WAL logging scheme >> needs to be redesigned. > > I'm not sure what a good fix is. I was thinking that maybe if a > placeholder tuple is found during index scan, and the corresponding > process is no longer running, then the index scanner would remove the > placeholder tuple, making the range unsummarized again. However, how > would we know that the process is gone? > > Another idea is to use WAL's rm_cleanup: during replay, remember that a > placeholder tuple was seen, then remove the info if we see an update > from the originating process that replaces the placeholder tuple with a > real one; at cleanup time, if the list of remembered placeholder tuples > is nonempty, remove them. > > (I vaguely recall we used the WAL rm_cleanup mechanism for something > like this, but we no longer do AFAICS.) > > ... Oh, but if there is a promotion involved, we may end up with a > placeholder insertion before the promotion and the update afterwards. > That would probably not be handled well. > Right, I think we'd miss those. And can't that happen for regular recovery too. If the placeholder tuple is before the redo LSN, we'd miss it too, right? But something prevents that. I think the root cause is the two WAL records are not linked together, so we fail to ensure the atomicity of the operation. Trying to fix this by tracking placeholder tuples seems like solving the symptoms. The easiest solution would be to link the records by XID, but of course that goes against the whole placeholder tuple idea - no one could modify the placeholder tuple in between. Maybe that's a price we have to pay. > > One thing not completely clear to me is whether this only affects > placeholder tuples. Is it possible to have this problem with regular > BRIN tuples? I think it isn't. > Yeah. I've been playing with this for a while, trying to cause issues with non-placeholder tuples. But I think that's fine - once the tuple becomes non-placeholder, all subsequent updates are part of the transaction that modifies the table. So if that fails, we don't update the index tuple. regards -- Tomas Vondra EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company