Thread: WAL replay should fdatasync() segments?
Hi, Currently, XLogInsert(), XLogFlush() or XLogBackgroundFlush() will write() data before fdatasync()ing them (duh, kinda obvious). But I think given the current recovery code that leaves a window where we can get into strange inconsistencies. Consider what happens if postgres (not the OS!) crashes after writing WAL data to the OS, but before fdatasync()ing it. Replay will happily read that record from disk and replay it, which is fine. At the end of recovery we then will start inserting new records, and those will be properly fsynced to disk. But if the *OS* crashes in that moment we might get into the strange situation where older records might be lost since they weren't fsync()ed, but newer records and the control file will persist. I think for a primary that window is relatively small, but I think it's a good bit bigger for a standby, especially if it's promoted. I think the correct way to handle this would be to fsync() segments we read from pg_xlog/ during recovery. Am I missing something? Greetings, Andres Freund -- Andres Freund http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training & Services
On Thu, Jan 23, 2014 at 1:21 AM, Andres Freund <andres@2ndquadrant.com> wrote: > Hi, > > Currently, XLogInsert(), XLogFlush() or XLogBackgroundFlush() will > write() data before fdatasync()ing them (duh, kinda obvious). But I > think given the current recovery code that leaves a window where we can > get into strange inconsistencies. > Consider what happens if postgres (not the OS!) crashes after writing > WAL data to the OS, but before fdatasync()ing it. Replay will happily > read that record from disk and replay it, which is fine. At the end of > recovery we then will start inserting new records, and those will be > properly fsynced to disk. > But if the *OS* crashes in that moment we might get into the strange > situation where older records might be lost since they weren't > fsync()ed, but newer records and the control file will persist. > > I think for a primary that window is relatively small, but I think it's > a good bit bigger for a standby, especially if it's promoted. In normal streaming replication case, ISTM that window is not bigger for the standby because basically the standby replays only the WAL data which walreceiver fsync'd to the disk. But if it replays the WAL file which was fetched from the archive, that WAL file might not have been flushed to the disk yet. In this case, that window might become bigger... Regards, -- Fujii Masao
On 2014-01-23 02:05:48 +0900, Fujii Masao wrote: > On Thu, Jan 23, 2014 at 1:21 AM, Andres Freund <andres@2ndquadrant.com> wrote: > > Hi, > > > > Currently, XLogInsert(), XLogFlush() or XLogBackgroundFlush() will > > write() data before fdatasync()ing them (duh, kinda obvious). But I > > think given the current recovery code that leaves a window where we can > > get into strange inconsistencies. > > Consider what happens if postgres (not the OS!) crashes after writing > > WAL data to the OS, but before fdatasync()ing it. Replay will happily > > read that record from disk and replay it, which is fine. At the end of > > recovery we then will start inserting new records, and those will be > > properly fsynced to disk. > > But if the *OS* crashes in that moment we might get into the strange > > situation where older records might be lost since they weren't > > fsync()ed, but newer records and the control file will persist. > > > > I think for a primary that window is relatively small, but I think it's > > a good bit bigger for a standby, especially if it's promoted. > > In normal streaming replication case, ISTM that window is not bigger for > the standby because basically the standby replays only the WAL data > which walreceiver fsync'd to the disk. But if it replays the WAL file which > was fetched from the archive, that WAL file might not have been flushed > to the disk yet. In this case, that window might become bigger... Yea, but if the walreceiver receives data and crashes/disconnects before fsync(), we'll read it from pg_xlog, rigth? And if we promote, we'll start inserting new records before establishing a new checkpoint. Greetings, Andres Freund -- Andres Freund http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training & Services
On Thu, Jan 23, 2014 at 2:08 AM, Andres Freund <andres@2ndquadrant.com> wrote: > On 2014-01-23 02:05:48 +0900, Fujii Masao wrote: >> On Thu, Jan 23, 2014 at 1:21 AM, Andres Freund <andres@2ndquadrant.com> wrote: >> > Hi, >> > >> > Currently, XLogInsert(), XLogFlush() or XLogBackgroundFlush() will >> > write() data before fdatasync()ing them (duh, kinda obvious). But I >> > think given the current recovery code that leaves a window where we can >> > get into strange inconsistencies. >> > Consider what happens if postgres (not the OS!) crashes after writing >> > WAL data to the OS, but before fdatasync()ing it. Replay will happily >> > read that record from disk and replay it, which is fine. At the end of >> > recovery we then will start inserting new records, and those will be >> > properly fsynced to disk. >> > But if the *OS* crashes in that moment we might get into the strange >> > situation where older records might be lost since they weren't >> > fsync()ed, but newer records and the control file will persist. >> > >> > I think for a primary that window is relatively small, but I think it's >> > a good bit bigger for a standby, especially if it's promoted. >> >> In normal streaming replication case, ISTM that window is not bigger for >> the standby because basically the standby replays only the WAL data >> which walreceiver fsync'd to the disk. But if it replays the WAL file which >> was fetched from the archive, that WAL file might not have been flushed >> to the disk yet. In this case, that window might become bigger... > > Yea, but if the walreceiver receives data and crashes/disconnects before > fsync(), we'll read it from pg_xlog, rigth? And if we promote, we'll > start inserting new records before establishing a new checkpoint. Yeah, true. Such unflushed WAL file can be read by the subsequent recovery... Regards, -- Fujii Masao