Re: BUGFIX: standby disconnect can corrupt serialized reorder buffers

On 06/03/18 09:37, Craig Ringer wrote:
>
> Revised patch attached.
>
> I have _not_ rewritten to use sscanf yet. I'll do that next, so you
> can choose the fewer-changes patch for backpatching if desired. 
>
>
> ... and I'm not convinced it's really an improvement.
>
>         uint32 xid, lsn_high, lsn_low;
>
>         if (sscanf(spill_de->d_name, REORDERBUFFER_TXN_FILENAME_FORMAT,
>                xid, lsn_high, lsn_low) == 3)
>         {
>
> since we don't use the scanned-out information.
>
> I guess my answer to causing problems if you create a file named
> "xidfoo" in a slot directory is yeah, don't do that.
>
> Note that this patch changes the PANIC to ERROR. This will promote to
> FATAL during the startup process, which I think is fine. Objections?
>
You mean because PG_exception_stack is not initialized for startup
process? That deserves comment I think.

Other than that if I am very nitpicky, I'd call the new function
ReorderBufferCleanupSerializedTXNs.

--
Petr Jelinek http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services