couchdb-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Adam Kocoloski <kocol...@apache.org>
Subject Re: optimal settings for [couchdb] fsync_options?
Date Wed, 14 Apr 2010 15:55:43 GMT
Bah, that's quite right.  Thanks for the step-by-step, I'm not sure how I missed it before.

Adam

On Apr 14, 2010, at 11:04 AM, Robert Newson wrote:

> I think Damien is right here. Consider this sequence;
> 
> 1) update btree
> 2) fsync
> 3) write new header
> 4) fsync
> 5) more updates
> 6) fsync
> 7) write new header
> 8) process terminates
> 
> On open, the header at 7) might or might not be flushed all the way to
> disk, but couchdb would update views to include changes made at 5).
> Since the header at 7) isn't definitely fsync'ed, a second crash (say,
> a kernel panic) could revert the .couch file itself to the state at
> 4), but views are permanently wrong. It's hard to see it in practice
> because the header is 4k and almost always gets to disk soon enough
> anyway, especially if you do more i/o on the view indexes.
> 
> B.
> 
> On Wed, Apr 14, 2010 at 3:46 PM, Adam Kocoloski <kocolosk@apache.org> wrote:
>> Thanks Damien.  I'm thinking that the situation you describe cannot occur if before_header
is enabled in the fsync_options, since any data pointed to by the #db_header that the server
found after the restart was already synced.  Is that correct?
>> 
>> Adam
>> 
>> On Apr 14, 2010, at 10:26 AM, Damien Katz wrote:
>> 
>>> The reason for fsync on open is the server doesn't know if the data it's reading
off the file is commited fully to the disk. It's possible the the server wrote to file and
crashed before fsync, then restarted. Then it could refresh view indexes on the non-fsynced
storage data, for example, and crash again, losing data in the storage file, but not the updates
to the index file. Now the index is permanently out of date with the storage file. But if
you fsync on opening the storage file, that can't happen.
>>> 
>>> -Damien
>>> 
>>> 
>>> On Apr 14, 2010, at 5:52 AM, Adam Kocoloski wrote:
>>> 
>>>> Initially posted on user@, but maybe it got lost in the noise.  Does anyone
know why we call fsync when we open a file?
>>>> 
>>>> Adam
>>>> 
>>>> Begin forwarded message:
>>>> 
>>>>> From: Adam Kocoloski <kocolosk@apache.org>
>>>>> Date: April 11, 2010 10:44:03 PM EDT
>>>>> To: user@couchdb.apache.org
>>>>> Subject: optimal settings for [couchdb] fsync_options?
>>>>> 
>>>>> Hi folks, I wanted to assemble some concrete information about the purpose
of each of the three fsync_options available in CouchDB and under what conditions they should
be enabled/disabled.  These options are
>>>>> 
>>>>> 1) before_header - calls file:sync(Fd) before writing a DB header to
disk.  I believe the goal here is to prevent DB corruption by ensuring that all the data referred
to by the header is durably stored before the header is written.  A system that preserves
write ordering could safely disable this option.  Does anyone know an example of such a system?
Perhaps a combination of a noop IO scheduler and a write-through or nonvolatile disk cache?
>>>>> 
>>>>> 2) after_header - calls file:sync(Fd) immediately after writing the DB
header.  I think this one is done so that we don't lose too much data following a CouchDB
restart, and so that a client can ensure that stored data will be retrievable after a restart
by POSTing to /db/_ensure_full_commit.  It might make sense to disable this option if e.g.
you're relying on replication for durability.  Although that's dicey because the replicator
calls ensure_full_commit for both DBs before writing its own checkpoint record*, and by disabling
the after_header option you'd run the risk of skipping updates on the target in the face of
a power failure.
>>>>> 
>>>>> 3) on_file_open - calls file:sync(Fd) immediately after opening a DB
file.  I really don't know the purpose of this one.  Anyone?
>>>>> 
>>>>> Best, Adam
>>>>> 
>>>>> * The reason the replicator calls ensure_full_commit on the source is
to detect situations where update_seqs might be reused.  I wonder if we could engineer a way
around that ever happening, for example by ensuring that on restart the update sequence jumps
by a large number.  But that's a discussion for dev@.
>>>> 
>>> 
>> 
>> 


Mime
View raw message