Re: WAL reader APIs and WAL segment open/close callbacks

Kyotaro Horiguchi <horikyota.ntt@gmail.com>

From: Kyotaro Horiguchi <horikyota.ntt@gmail.com>
To: michael@paquier.xyz
Cc: pgsql-hackers@lists.postgresql.org, ah@cybertec.at, alvherre@2ndquadrant.com
Date: 2020-05-25T02:17:06Z
Lists: pgsql-hackers
At Mon, 25 May 2020 07:44:09 +0900, Michael Paquier <michael@paquier.xyz> wrote in 
> Hi all,
> 
> I have been playing with the new APIs of xlogreader.h, and while
> merging some of my stuff with 13, I found the handling around
> ->seg.ws_file overcomplicated and confusing as it is necessary for a
> plugin to manipulate directly the fd of an opened segment in the WAL
> segment open/close callbacks.

That depends on where we draw responsibility border, or who is
responsible to the value of ws_file.  I think that this API change was
assuming the callbacks having full-knowledge of the xlogreader struct
and are responsible to maintain related struct members, and I agree to
that direction.

> Wouldn't it be cleaner to limit the exposition of ->seg.ws_file to the
> user if possible?  There are cases like a WAL sender where you cannot
> do that, but something that came to my mind is to make
> WALSegmentOpenCB return the fd of the opened segment, and pass down the
> fd to close to WALSegmentCloseCB.  Then xlogreader.c is in charge of
> resetting the field when a segment is closed.
> 
> Any thoughts?

If we are going to hide the struct from the callbacks, we shouldn't
pass to the callbacks a pointer to the complete XLogReaderState
struct.

regards.

-- 
Kyotaro Horiguchi
NTT Open Source Software Center



Commits

  1. Fix some comments in xlogreader.h

  2. Adjust walsender usage of xlogreader, simplify APIs