You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@couchdb.apache.org by "Adam Kocoloski (JIRA)" <ji...@apache.org> on 2010/06/07 18:37:43 UTC

[jira] Commented: (COUCHDB-786) Make fewer gen_server calls to fsync the db header

    [ https://issues.apache.org/jira/browse/COUCHDB-786?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12876305#action_12876305 ] 

Adam Kocoloski commented on COUCHDB-786:
----------------------------------------

I'm worried about the interaction of this patch with COUCHDB-767.  That issue contains a patch which causes couch_file to spawn a worker process to sync the file.  The idea behind COUCHDB-767 is that reads can proceed even while the file is being synced.

It seems that this patch would have the opposite effect -- all readers would be blocked, not just while a file is synced, but while the header is being written (which typically involves 2 syncs).

Perhaps it would be a good idea to combine the strategies from both issues, and spawn a function to do the sync-write-sync dance for write_header?  The challenge is that #file.eof needs to be updated when the header is written, which is not the case for a simple sync.

> Make fewer gen_server calls to fsync the db header
> --------------------------------------------------
>
>                 Key: COUCHDB-786
>                 URL: https://issues.apache.org/jira/browse/COUCHDB-786
>             Project: CouchDB
>          Issue Type: Improvement
>          Components: Database Core
>    Affects Versions: 0.10.2, 0.11
>            Reporter: Randall Leeds
>             Fix For: 0.12
>
>         Attachments: 0001-db-header-fsync-with-fewer-gen_server-calls.patch
>
>
> Currently the couch_db_updater process is responsible for calling couch_file:sync based on the fsync options when the db header is written. Instead, this patch just passes the fsync options down to couch_file so that writing the header can always be performed with a single gen_server call.
> Writes are serialized through couch_db_updater but reads can have high concurrency. This patch should boost write performance when the message queue on couch_file is long.

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.