Hi,
On 2017-04-24 21:17:43 -0700, Andres Freund wrote:
> I've lately seen more and more installations where the generation of
> write-ahead-log (WAL) is one of the primary bottlenecks. I'm curious
> whether that's primarily a "sampling error" of mine, or whether that's
> indeed more common.
>
> The primary reason I'm curious is that I'm pondering a few potential
> optimizations, and would like to have some guidance which are more and
> which are less important.
>
> Questions (answer as many you can comfortably answer):
> - How many MB/s, segments/s do you see on busier servers?
> - What generates the bulk of WAL on your servers (9.5+ can use
> pg_xlogdump --stats to compute that)?
> - Are you seeing WAL writes being a bottleneck?OA
> - What kind of backup methods are you using and is the WAL volume a
> problem?
> - What kind of replication are you using and is the WAL volume a
> problem?
> - What are your settings for wal_compression, max_wal_size (9.5+) /
> checkpoint_segments (< 9.5), checkpoint_timeout and wal_buffers?
> - Could you quickly describe your workload?
Ok, based on the, few, answers I've got so far, my experience is indeed
skewed. A number of the PG users I interacted with over the last couple
years had WAL write ranges somewhere in the range of 500MB/s to 2.2GB/s
(max I'veseen). At that point WAL insertion became a major bottleneck,
even if storage was more than fast enough to keep up. To address these
we'd need some changes, but the feedback so far suggest that it's not
yet a widespread issue...
- Andres