Обсуждение: High checkpoint_segments
We need to do a few bulk updates as Rails migrations. We're a typical read-mostly web site, so at the moment, our checkpoint settings and WAL are all default (3 segments, 5 min, 16MB), and updating a million rows takes 10 minutes due to all the checkpointing. We have no replication or hot standbys. As a consumer-web startup, with no SLA, and not a huge database, and if we ever do have to recover from downtime it's ok if it takes longer.. is there a reason NOT to always run with something like checkpoint_segments = 1000, as long as I leave the timeout at 5m? Jay Levitt
On Wed, Feb 15, 2012 at 1:35 AM, Jay Levitt <jay.levitt@gmail.com> wrote:
We need to do a few bulk updates as Rails migrations. We're a typical read-mostly web site, so at the moment, our checkpoint settings and WAL are all default (3 segments, 5 min, 16MB), and updating a million rows takes 10 minutes due to all the checkpointing.
We have no replication or hot standbys. As a consumer-web startup, with no SLA, and not a huge database, and if we ever do have to recover from downtime it's ok if it takes longer.. is there a reason NOT to always run with something like checkpoint_segments = 1000, as long as I leave the timeout at 5m?
16MB * 1000 = 16000MB worth pg_xlog data, which is not advisable from I/O perspective and data loss perspective. Even in the most unimaginable case if all of these 1000 files get filled up in less than 5 mins, there are chances that system will slow down due to high IO and CPU.
You may think of increasing checkpoint_timeout as well, but, some monitoring and analysis is needed to arrive at a number.
What does pg_stat_bgwriter say about checkpoints ?
Do you have log_checkpoints enabled ?
Thanks
VB
On Tue, Feb 14, 2012 at 10:57 PM, Venkat Balaji <venkat.balaji@verse.in> wrote: > > On Wed, Feb 15, 2012 at 1:35 AM, Jay Levitt <jay.levitt@gmail.com> wrote: >> >> We need to do a few bulk updates as Rails migrations. We're a typical >> read-mostly web site, so at the moment, our checkpoint settings and WAL are >> all default (3 segments, 5 min, 16MB), and updating a million rows takes 10 >> minutes due to all the checkpointing. >> >> We have no replication or hot standbys. As a consumer-web startup, with >> no SLA, and not a huge database, and if we ever do have to recover from >> downtime it's ok if it takes longer.. is there a reason NOT to always run >> with something like checkpoint_segments = 1000, as long as I leave the >> timeout at 5m? > > > Still checkpoints keep occurring every 5 mins. Anyways > checkpoint_segments=1000 is huge, this implies you are talking about > 16MB * 1000 = 16000MB worth pg_xlog data, which is not advisable from I/O > perspective and data loss perspective. Even in the most unimaginable case if > all of these 1000 files get filled up in less than 5 mins, there are chances > that system will slow down due to high IO and CPU. As far as I know there is no data loss issue with a lot of checkpoint segments.
Jay Levitt wrote: > We need to do a few bulk updates as Rails migrations. We're a typical > read-mostly web site, so at the moment, our checkpoint settings and WAL are > all default (3 segments, 5 min, 16MB), and updating a million rows takes 10 > minutes due to all the checkpointing. > > We have no replication or hot standbys. As a consumer-web startup, with no > SLA, and not a huge database, and if we ever do have to recover from > downtime it's ok if it takes longer.. is there a reason NOT to always run > with something like checkpoint_segments = 1000, as long as I leave the > timeout at 5m? There's nothing wrong with the idea except for the amount of WAL and a huge checkpoint that can stall your system for a while in a worst-case scenario. You can't get rid of checkpoint I/O completely. I'd tune to a more conservative value, maybe 30 or at most 100 and see if that solves your problem. Check statistics to see if checkpoints are time-driven or not. As soon as almost all checkpoints are time- driven, further raising of checkpoint_segments won't do anything for you. Yours, Laurenz Albe
On Wed, Feb 15, 2012 at 12:21 PM, Scott Marlowe <scott.marlowe@gmail.com> wrote:
On Tue, Feb 14, 2012 at 10:57 PM, Venkat Balaji <venkat.balaji@verse.in> wrote:
>
> On Wed, Feb 15, 2012 at 1:35 AM, Jay Levitt <jay.levitt@gmail.com> wrote:
>>
>> We need to do a few bulk updates as Rails migrations. We're a typical
>> read-mostly web site, so at the moment, our checkpoint settings and WAL are
>> all default (3 segments, 5 min, 16MB), and updating a million rows takes 10
>> minutes due to all the checkpointing.
>>
>> We have no replication or hot standbys. As a consumer-web startup, with
>> no SLA, and not a huge database, and if we ever do have to recover from
>> downtime it's ok if it takes longer.. is there a reason NOT to always run
>> with something like checkpoint_segments = 1000, as long as I leave the
>> timeout at 5m?
>
>
> Still checkpoints keep occurring every 5 mins. Anyways
> checkpoint_segments=1000 is huge, this implies you are talking about
> 16MB * 1000 = 16000MB worth pg_xlog data, which is not advisable from I/O
> perspective and data loss perspective. Even in the most unimaginable case if
> all of these 1000 files get filled up in less than 5 mins, there are chances
> that system will slow down due to high IO and CPU.
As far as I know there is no data loss issue with a lot of checkpoint segments.
Data loss would be an issue when there is a server crash or pg_xlog crash etc. That many number of pg_xlog files (1000) would contribute to huge data loss (data changes not synced to the base are not guaranteed). Of-course, this is not related to the current situation. Normally we calculate the checkpoint completion time, IO pressure, CPU load and the threat to the data loss when we configure checkpoint_segments.
On 15 Únor 2012, 10:38, Venkat Balaji wrote: > > Data loss would be an issue when there is a server crash or pg_xlog crash > etc. That many number of pg_xlog files (1000) would contribute to huge > data > loss (data changes not synced to the base are not guaranteed). Of-course, > this is not related to the current situation. Normally we calculate the > checkpoint completion time, IO pressure, CPU load and the threat to the > data loss when we configure checkpoint_segments. So you're saying that by using small number of checkpoint segments you limit the data loss when the WAL gets corrupted/lost? That's a bit like buying a Maseratti and then not going faster than 10mph because you might crash at higher speeds ... The problem here is that the WAL is usually placed on more reliable drives (compared to the data files) or a RAID1 array and as it's just writing data sequentially, so the usage pattern is much less likely to cause data/drive corruption (compared to data files that need to handle a lot of random I/O, etc.). So while it possible the WAL might get corrupted, the probability of data file corruption is much higher. And the corruption might easily happen silently during a checkpoint, so there won't be any WAL segments no matter how many of them you keep ... And by using low number of checkpoint segments it actually gets worse, because it means more frequent checkpoints -> more I/O on the drives -> more wearout of the drives etc. If you need to protect yourself against this, you need to keep a WAL archive (prefferably on a separate machine) and/or a hot standby for failover. kind regards Tomas
On Wednesday, February 15, 2012 10:38:23 AM Venkat Balaji wrote: > On Wed, Feb 15, 2012 at 12:21 PM, Scott Marlowe <scott.marlowe@gmail.com>wrote: > > On Tue, Feb 14, 2012 at 10:57 PM, Venkat Balaji <venkat.balaji@verse.in> > > > all of these 1000 files get filled up in less than 5 mins, there are > > > chances that system will slow down due to high IO and CPU. > > As far as I know there is no data loss issue with a lot of checkpoint > > segments. > Data loss would be an issue when there is a server crash or pg_xlog crash > etc. That many number of pg_xlog files (1000) would contribute to huge data > loss (data changes not synced to the base are not guaranteed). Of-course, > this is not related to the current situation. Normally we calculate the > checkpoint completion time, IO pressure, CPU load and the threat to the > data loss when we configure checkpoint_segments. I think you might be misunderstanding something. A high number of checkpoint_segments can lead to slower recovery - all those changes need to be reapplied - but it won't lead to lost data. The data inside the wal will be fsynced at appropriate times (commit; background writer; too much written). Andres
> Data loss would be an issue when there is a server crash or pg_xlog crashSo you're saying that by using small number of checkpoint segments you
> etc. That many number of pg_xlog files (1000) would contribute to huge
> data
> loss (data changes not synced to the base are not guaranteed). Of-course,
> this is not related to the current situation. Normally we calculate the
> checkpoint completion time, IO pressure, CPU load and the threat to the
> data loss when we configure checkpoint_segments.
limit the data loss when the WAL gets corrupted/lost? That's a bit like
buying a Maseratti and then not going faster than 10mph because you might
crash at higher speeds ...
No. I am not saying that checkpoint_segments must be lower. I was just trying to explain the IO over-head on putting high (as high as 1000) checkpoint segments. Lower number of checkpoint segments will lead to more frequent IOs which is not good. Agreed.
The problem here is that the WAL is usually placed on more reliable drives
(compared to the data files) or a RAID1 array and as it's just writing
data sequentially, so the usage pattern is much less likely to cause
data/drive corruption (compared to data files that need to handle a lot of
random I/O, etc.).
Agreed.
So while it possible the WAL might get corrupted, the probability of data
file corruption is much higher. And the corruption might easily happen
silently during a checkpoint, so there won't be any WAL segments no matter
how many of them you keep ...
Agreed. When corruption occurs, it really does not matter how many WAL segments are kept in pg_xlog.
But, at any point of time if PG needs
And by using low number of checkpoint segments it actually gets worse,
because it means more frequent checkpoints -> more I/O on the drives ->
more wearout of the drives etc.
Completely agreed. As mentioned above. I choose checkpoint_segments and checkpoint_timeout once i observe the checkpoint behavior.
If you need to protect yourself against this, you need to keep a WAL
archive (prefferably on a separate machine) and/or a hot standby for
failover.
WAL archiving is a different situation where-in you need to backup the pg_xlog files by enabling archiving.
I was referring to an exclusive situation, where-in pg_xlogs are not archived and data is not yet been synced to base files (by bgwriter) and the system crashed, then PG would depend on pg_xlog to recover and reach the consistent state, if the pg_xlog is also not available, then there would be a data loss and this depends on how much data is present in pg_xlog files.
Thanks,
VB
On Wed, Feb 15, 2012 at 4:12 PM, Andres Freund <andres@anarazel.de> wrote:
On Wednesday, February 15, 2012 10:38:23 AM Venkat Balaji wrote:
> On Wed, Feb 15, 2012 at 12:21 PM, Scott Marlowe
<scott.marlowe@gmail.com>wrote:
> > On Tue, Feb 14, 2012 at 10:57 PM, Venkat Balaji <venkat.balaji@verse.in>> > > all of these 1000 files get filled up in less than 5 mins, there areI think you might be misunderstanding something. A high number of
> > > chances that system will slow down due to high IO and CPU.
> > As far as I know there is no data loss issue with a lot of checkpoint
> > segments.
> Data loss would be an issue when there is a server crash or pg_xlog crash
> etc. That many number of pg_xlog files (1000) would contribute to huge data
> loss (data changes not synced to the base are not guaranteed). Of-course,
> this is not related to the current situation. Normally we calculate the
> checkpoint completion time, IO pressure, CPU load and the threat to the
> data loss when we configure checkpoint_segments.
checkpoint_segments can lead to slower recovery - all those changes need to be
reapplied - but it won't lead to lost data. The data inside the wal will be
fsynced at appropriate times (commit; background writer; too much written).
Recovery would take time because all the changes in WAL files of pg_xlog (which is high) must be replayed to reach consistent state. When disaster strikes and if pg_xlogs are not available and data in WAL is not fsynced yet, then recovery is not possible and data loss will be huge. It also depends on how much data is not fsynced.
Thanks,
VB