Thread: [GENERAL] Strange case of database bloat
Hi;
--
First, I haven't seen major problems of database bloat in a long time which is why I find this case strange. I wanted to ask here what may be causing it.
Problem:
==========
Database is in the 100GB to 200GB size range, running on btrfs (not my choice) with nodatacow enabled (which I set up to fix a performance issue). The workload is a very heavy batch-update workload.
The database bloats linearly. I have measured this on one table (of 149M rows).
After vacuum full this table is (including indexes): 17GB
Every 24 hrs, seems to add its original space in size to the file system +/-.
Bloat seems to be affecting both indexes and underlying tables.
Vacuum verbose does not indicate a disproportionate number of rows being unremovable. So autovacuum is keeping up without too much difficulty.
Troubleshooting so far
=======================
filefrag finds a single extent on each file, so copy-on-write is not the culprit
Selecting the smallest 10 values of ctid from one of the bloating tables shows the first page used is around page 35 with one row per used page (and large gaps in between).
Questions
===========
I assume that it is the fact that rows update frequently which is the problem here? But why doesn't Postgres re-use any of the empty disk pages?
More importantly, is there anything that can be done to mitigate this issue other than a frequent vacuum full?
Best Wishes,
Chris Travers
Efficito: Hosted Accounting and ERP. Robust and Flexible. No vendor lock-in.
On Wed, Jul 5, 2017 at 7:18 AM, Chris Travers <chris.travers@gmail.com> wrote:
Hi;First, I haven't seen major problems of database bloat in a long time which is why I find this case strange. I wanted to ask here what may be causing it.Problem:==========Database is in the 100GB to 200GB size range, running on btrfs (not my choice) with nodatacow enabled (which I set up to fix a performance issue). The workload is a very heavy batch-update workload.The database bloats linearly. I have measured this on one table (of 149M rows).After vacuum full this table is (including indexes): 17GBEvery 24 hrs, seems to add its original space in size to the file system +/-.Bloat seems to be affecting both indexes and underlying tables.Vacuum verbose does not indicate a disproportionate number of rows being unremovable. So autovacuum is keeping up without too much difficulty.Troubleshooting so far=======================filefrag finds a single extent on each file, so copy-on-write is not the culpritSelecting the smallest 10 values of ctid from one of the bloating tables shows the first page used is around page 35 with one row per used page (and large gaps in between).Questions===========I assume that it is the fact that rows update frequently which is the problem here? But why doesn't Postgres re-use any of the empty disk pages?More importantly, is there anything that can be done to mitigate this issue other than a frequent vacuum full?
Two points I think I forgot to mention:
This is PostgreSQL 9.5.1
Last I saw something similar was a more "minor" case on a larger db, on PostgreSQL 9.3.x
The more minor case was a small table (maybe 20k rows) which had bloated to 1GB in size due to this same sort of problem but we ignored it because the table was cached all the time and at the RAM we were using, it wasn't a significant drain on performance. However, here it is.
First 20 CTIDs from one table:
(35,25)
(48,15)
(76,20)
(77,20)
(83,20)
(96,19)
(100,19)
(103,13)
(111,9)
(115,12)
(124,11)
(120,12)
(131,12)
(137,12)
(150,14)
(152,12)
(157,20)
(162,14)
--Best Wishes,Chris TraversEfficito: Hosted Accounting and ERP. Robust and Flexible. No vendor lock-in.
Best Wishes,
Chris Travers
Efficito: Hosted Accounting and ERP. Robust and Flexible. No vendor lock-in.
On Wed, 5 Jul 2017 07:18:03 +0200 Chris Travers <chris.travers@gmail.com> wrote: > Hi; > > First, I haven't seen major problems of database bloat in a long time which > is why I find this case strange. I wanted to ask here what may be causing > it. > > Problem: > ========== > Database is in the 100GB to 200GB size range, running on btrfs (not my > choice) with nodatacow enabled (which I set up to fix a performance > issue). The workload is a very heavy batch-update workload. > > The database bloats linearly. I have measured this on one table (of 149M > rows). > > After vacuum full this table is (including indexes): 17GB > Every 24 hrs, seems to add its original space in size to the file system > +/-. > > Bloat seems to be affecting both indexes and underlying tables. > > Vacuum verbose does not indicate a disproportionate number of rows being > unremovable. So autovacuum is keeping up without too much difficulty. > > > Troubleshooting so far > ======================= > > filefrag finds a single extent on each file, so copy-on-write is not the > culprit > > Selecting the smallest 10 values of ctid from one of the bloating tables > shows the first page used is around page 35 with one row per used page (and > large gaps in between). > > Questions > =========== > I assume that it is the fact that rows update frequently which is the > problem here? But why doesn't Postgres re-use any of the empty disk pages? > > More importantly, is there anything that can be done to mitigate this issue > other than a frequent vacuum full? 2x the working size for a frequently updated table isn't terrible bloat. Or are you saying it grows 2x every 24 hours and keeps growing? The real question is how often the table is being vacuumed. How long have you let the experiment run for? Does the table find an equilibrium size where it stops growing? Have you turned on logging for autovacuum to see how often it actually runs on this table? No unremovable rows does not indicate that autovaccum is keeping up. It just indicates that you don't have a problem with uncommitted transactions holding rows for long periods of time. Have you looked at tuning the autovacuum parameters for this table? More frequent vacuums should keep things more under control. However, if the write load is heavy, you'll probably want to lower autovacuum_vacuum_cost_delay. Personally, I feel like the default value for this should be 0, but there are likely those that would debate that. In any event, if that setting is too high it can cause autovacuum to take so long that it can't keep up. In theory, setting it too low can cause autovaccum to have a negative performance impact, but I've never seen that happen on modern hardware. But that's all speculation until you know how frequently autovacuum runs on that table and how long it takes to do its work. -- PT <wmoran@potentialtech.com>
On Wed, Jul 5, 2017 at 1:00 PM, PT <wmoran@potentialtech.com> wrote:
2x the working size for a frequently updated table isn't terrible bloat. Or are
you saying it grows 2x every 24 hours and keeps growing? The real question is
how often the table is being vacuumed. How long have you let the experiment run
for? Does the table find an equilibrium size where it stops growing? Have you
turned on logging for autovacuum to see how often it actually runs on this
table?
If it were only twice it would not bother me. The fact that it is twice after 24 hrs, 3x after 48 hrs and 4x after 72 hrs is alarming.
No unremovable rows does not indicate that autovaccum is keeping up. It just
indicates that you don't have a problem with uncommitted transactions holding
rows for long periods of time.
Right. I should have specified that I also have not seen auto vacuum in pg_stat_activity with an unusual duration.
Have you looked at tuning the autovacuum parameters for this table? More frequent
vacuums should keep things more under control. However, if the write load is
heavy, you'll probably want to lower autovacuum_vacuum_cost_delay. Personally,
I feel like the default value for this should be 0, but there are likely those
that would debate that. In any event, if that setting is too high it can cause
autovacuum to take so long that it can't keep up. In theory, setting it too low
can cause autovaccum to have a negative performance impact, but I've never seen
that happen on modern hardware.
Most of the writes are periodic (hourly?) batch updates which are fairly big.
But that's all speculation until you know how frequently autovacuum runs on
that table and how long it takes to do its work.
Given the other time I have seen similar behaviour, the question in my mind is why free pages near the beginning of the table don't seem to be re-used.
I would like to try to verify that however, if you have any ideas.
--
PT <wmoran@potentialtech.com>
Best Wishes,
Chris Travers
Efficito: Hosted Accounting and ERP. Robust and Flexible. No vendor lock-in.
On Wed, Jul 5, 2017 at 7:28 AM, Chris Travers <chris.travers@gmail.com> wrote:
On Wed, Jul 5, 2017 at 1:00 PM, PT <wmoran@potentialtech.com> wrote:2x the working size for a frequently updated table isn't terrible bloat. Or are
you saying it grows 2x every 24 hours and keeps growing? The real question is
how often the table is being vacuumed. How long have you let the experiment run
for? Does the table find an equilibrium size where it stops growing? Have you
turned on logging for autovacuum to see how often it actually runs on this
table?If it were only twice it would not bother me. The fact that it is twice after 24 hrs, 3x after 48 hrs and 4x after 72 hrs is alarming.
No unremovable rows does not indicate that autovaccum is keeping up. It just
indicates that you don't have a problem with uncommitted transactions holding
rows for long periods of time.Right. I should have specified that I also have not seen auto vacuum in pg_stat_activity with an unusual duration.
What about anything 'WHERE state = 'idle in transaction' ?
Have you looked at tuning the autovacuum parameters for this table? More frequent
vacuums should keep things more under control. However, if the write load is
heavy, you'll probably want to lower autovacuum_vacuum_cost_delay. Personally,
I feel like the default value for this should be 0, but there are likely those
that would debate that. In any event, if that setting is too high it can cause
autovacuum to take so long that it can't keep up. In theory, setting it too low
can cause autovaccum to have a negative performance impact, but I've never seen
that happen on modern hardware.Most of the writes are periodic (hourly?) batch updates which are fairly big.
I've had similar issues when each update makes a row larger than any of the available slots. I had a workload (admittedly on an older version of postgres) where we were updating every row a few times a day. Each time, the row (a bytea field) would grow about 0.5 - 5.0%. This would prevent us from using freespace (it was all too small). The only way around this was :
1. Run manual table rebuilds (this was before pg_repack / reorg). Use pg_repack now
2. Fix the app
Essentially, I would do targeted, aggressive vacuuming and then, once a month (or once I hit a bloat threshold) do a repack (again, it was my custom process back then). This was the bandage until I could get the app fixed to stop churning so badly.
But that's all speculation until you know how frequently autovacuum runs on
that table and how long it takes to do its work.Given the other time I have seen similar behaviour, the question in my mind is why free pages near the beginning of the table don't seem to be re-used.I would like to try to verify that however, if you have any ideas.
--
PT <wmoran@potentialtech.com>--Best Wishes,Chris TraversEfficito: Hosted Accounting and ERP. Robust and Flexible. No vendor lock-in.
On Wed, 5 Jul 2017 13:28:29 +0200 Chris Travers <chris.travers@gmail.com> wrote: > On Wed, Jul 5, 2017 at 1:00 PM, PT <wmoran@potentialtech.com> wrote: > > > 2x the working size for a frequently updated table isn't terrible bloat. > > Or are > > you saying it grows 2x every 24 hours and keeps growing? The real question > > is > > how often the table is being vacuumed. How long have you let the > > experiment run > > for? Does the table find an equilibrium size where it stops growing? Have > > you > > turned on logging for autovacuum to see how often it actually runs on this > > table? > > If it were only twice it would not bother me. The fact that it is twice > after 24 hrs, 3x after 48 hrs and 4x after 72 hrs is alarming. Ok, yup, that seems like an issue. > > No unremovable rows does not indicate that autovaccum is keeping up. It > > just > > indicates that you don't have a problem with uncommitted transactions > > holding > > rows for long periods of time. > > Right. I should have specified that I also have not seen auto vacuum in > pg_stat_activity with an unusual duration. How long does it take when you run it manually? My experience is that autovac can take orders of magnitude longer with the default cost delays, but just becuase you don't see it, doesn't mean it's not happening. Turn on autovac logging and check the logs after a few days. > > Have you looked at tuning the autovacuum parameters for this table? More > > frequent > > vacuums should keep things more under control. However, if the write load > > is > > heavy, you'll probably want to lower autovacuum_vacuum_cost_delay. > > Personally, > > I feel like the default value for this should be 0, but there are likely > > those > > that would debate that. In any event, if that setting is too high it can > > cause > > autovacuum to take so long that it can't keep up. In theory, setting it > > too low > > can cause autovaccum to have a negative performance impact, but I've never > > seen > > that happen on modern hardware. > > Most of the writes are periodic (hourly?) batch updates which are fairly > big. Not sure how that statement is related to the comments I made preceeding it. > > But that's all speculation until you know how frequently autovacuum runs on > > that table and how long it takes to do its work. > > Given the other time I have seen similar behaviour, the question in my mind > is why free pages near the beginning of the table don't seem to be re-used. It's possible that the early pages don't have enough usable space for the updated rows. Depending on your update patterns, you may end up with bloat scattered across many pages, with no individual page having enough space to be reused. That seems unlikely as the bloat becomes many times the used space, though. The pg_freespacemap extension should be useful in determining if that's what's happening. Combine that with turning on logging to ensure that autovacuum is actually operating effectively. -- Bill Moran <wmoran@potentialtech.com>
On Wed, Jul 5, 2017 at 3:51 PM, Bill Moran <wmoran@potentialtech.com> wrote:
On Wed, 5 Jul 2017 13:28:29 +0200
Chris Travers <chris.travers@gmail.com> wrote:
> On Wed, Jul 5, 2017 at 1:00 PM, PT <wmoran@potentialtech.com> wrote:
>
> > 2x the working size for a frequently updated table isn't terrible bloat.
> > Or are
> > you saying it grows 2x every 24 hours and keeps growing? The real question
> > is
> > how often the table is being vacuumed. How long have you let the
> > experiment run
> > for? Does the table find an equilibrium size where it stops growing? Have
> > you
> > turned on logging for autovacuum to see how often it actually runs on this
> > table?
>
> If it were only twice it would not bother me. The fact that it is twice
> after 24 hrs, 3x after 48 hrs and 4x after 72 hrs is alarming.
Ok, yup, that seems like an issue.
> > No unremovable rows does not indicate that autovaccum is keeping up. It
> > just
> > indicates that you don't have a problem with uncommitted transactions
> > holding
> > rows for long periods of time.
>
> Right. I should have specified that I also have not seen auto vacuum in
> pg_stat_activity with an unusual duration.
How long does it take when you run it manually? My experience is that autovac
can take orders of magnitude longer with the default cost delays, but just
becuase you don't see it, doesn't mean it's not happening. Turn on autovac
logging and check the logs after a few days.
a few min for a normal vacuum, maybe 20-30 min for vacuum full (on one of the large tables).
> > Have you looked at tuning the autovacuum parameters for this table? More
> > frequent
> > vacuums should keep things more under control. However, if the write load
> > is
> > heavy, you'll probably want to lower autovacuum_vacuum_cost_delay.
> > Personally,
> > I feel like the default value for this should be 0, but there are likely
> > those
> > that would debate that. In any event, if that setting is too high it can
> > cause
> > autovacuum to take so long that it can't keep up. In theory, setting it
> > too low
> > can cause autovaccum to have a negative performance impact, but I've never
> > seen
> > that happen on modern hardware.
>
> Most of the writes are periodic (hourly?) batch updates which are fairly
> big.
Not sure how that statement is related to the comments I made preceeding it.
Not using cost-based vacuum afaik but will check that. It is a good point.
> > But that's all speculation until you know how frequently autovacuum runs on
> > that table and how long it takes to do its work.
>
> Given the other time I have seen similar behaviour, the question in my mind
> is why free pages near the beginning of the table don't seem to be re-used.
It's possible that the early pages don't have enough usable space for the updated
rows. Depending on your update patterns, you may end up with bloat scattered across
many pages, with no individual page having enough space to be reused. That seems
unlikely as the bloat becomes many times the used space, though.
The fire 35 pages are completely empty. As I say I have seen this sort of thing before (and I wonder if empty pages early in a table are somehow biased against in terms of writes).
The pg_freespacemap extension should be useful in determining if that's what's
happening. Combine that with turning on logging to ensure that autovacuum is
actually operating effectively.
I am not convinced it is a vacuum problem. Would it be likely that large batch updates would linearly continue to write pages forward as a performance optimisation?
--
Bill Moran <wmoran@potentialtech.com>
Best Wishes,
Chris Travers
Efficito: Hosted Accounting and ERP. Robust and Flexible. No vendor lock-in.
On Tue, Jul 4, 2017 at 10:18 PM, Chris Travers <chris.travers@gmail.com> wrote:
Questions===========I assume that it is the fact that rows update frequently which is the problem here? But why doesn't Postgres re-use any of the empty disk pages?
Can you install the contrib extension pg_freespacemap and use "select * from pg_freespace('table_name')" to see if PostgreSQL agrees that the space is re-usable?
Cheers,
Jeff
On Tue, Jul 4, 2017 at 10:18 PM, Chris Travers <chris.travers@gmail.com> wrote: > First, I haven't seen major problems of database bloat in a long time which > is why I find this case strange. I wanted to ask here what may be causing > it. > > Problem: > ========== > Database is in the 100GB to 200GB size range, running on btrfs (not my > choice) with nodatacow enabled (which I set up to fix a performance issue). > The workload is a very heavy batch-update workload. > > The database bloats linearly. I have measured this on one table (of 149M > rows). > > After vacuum full this table is (including indexes): 17GB > Every 24 hrs, seems to add its original space in size to the file system > +/-. > > Bloat seems to be affecting both indexes and underlying tables. > > Vacuum verbose does not indicate a disproportionate number of rows being > unremovable. So autovacuum is keeping up without too much difficulty. That is odd. I find the linear increase in bloat alarming. I found this thread in the context of investigating a possible performance regression in 9.5 [1]. It's far from clear that that's actually what's up here, but it's worth considering as a possibility. Some questions about that: Can you show the schema of at least one of the tables that are most affected, its indexes, etc? Are all indexes bloated equally? Do you notice that unique indexes are not as badly affected as other indexes? How does it break down, in terms of how much each individual index grows in size? You say that the problem is with both indexes and tables. How much of this is table bloat, and how much is index bloat? Thanks [1] https://postgr.es/m/CAH2-Wz=SfAKVMv1x9Jh19EJ8am8TZn9f-yECipS9HrrRqSswnA@mail.gmail.com -- Peter Geoghegan