On Wed, 2021-01-06 at 10:19 -0600, Don Seiler wrote:
> This week we've noticed that we're starting to see spikes where COMMITs are
> taking much longer than usual.
> Sometimes, quite a few seconds to finish.
>
> This is PostgreSQL 12.4 on Ubuntu 18.04, all running in MS Azure (*not*
> managed by
Azure VMs do have their own IOPS limits that increase with increasing VM
"size". In this current case our VM size puts that VM IOPS limit well above
anything the disks are rated at, so it shouldn't be a bottleneck.
On Wed, Jan 6, 2021 at 1:15 PM Kenneth Marshall wrote:
> On Wed, Jan 06, 2021 at
On Wed, Jan 06, 2021 at 12:06:27PM -0600, Don Seiler wrote:
> On Wed, Jan 6, 2021 at 10:51 AM Joshua Drake wrote:
>
> Looking at the Azure portal metric, we are nowhere close to the advertised
> maximum IOPS or MB/s throughput (under half of the maximum IOPS and under a
> quarter of the MB/s maxi
>
> Looking at the Azure portal metric, we are nowhere close to the advertised
> maximum IOPS or MB/s throughput (under half of the maximum IOPS and under a
> quarter of the MB/s maximum). So there must be some other bottleneck in
> play. The IOPS limit on this VM size is even higher so that should
On Wed, Jan 6, 2021 at 10:51 AM Joshua Drake wrote:
> I have the stats_temp_directory in a tmpfs mount. I *do* have pg_wal on
>> the same premium SSD storage volume as the data directory. Normally I would
>> know to separate these but I was told with the cloud storage that it's all
>> virtualized
> I have the stats_temp_directory in a tmpfs mount. I *do* have pg_wal on
> the same premium SSD storage volume as the data directory. Normally I would
> know to separate these but I was told with the cloud storage that it's all
> virtualized anyway, plus storage IOPS are determined by disk size so
Good morning,
This week we've noticed that we're starting to see spikes where COMMITs are
taking much longer than usual. Sometimes, quite a few seconds to finish.
After a few minutes they disappear but then return seemingly at random.
This becomes visible to the app and end user as a big stall in
