Re: autovac issue with large number of tables - Mailing list pgsql-hackers

On 2020/12/01 16:23, Masahiko Sawada wrote:
> On Tue, Dec 1, 2020 at 1:48 PM Kasahara Tatsuhito
> <kasahara.tatsuhito@gmail.com> wrote:
>>
>> Hi,
>>
>> On Mon, Nov 30, 2020 at 8:59 PM Fujii Masao <masao.fujii@oss.nttdata.com> wrote:
>>>
>>>
>>>
>>> On 2020/11/30 10:43, Masahiko Sawada wrote:
>>>> On Sun, Nov 29, 2020 at 10:34 PM Kasahara Tatsuhito
>>>> <kasahara.tatsuhito@gmail.com> wrote:
>>>>>
>>>>> Hi, Thanks for you comments.
>>>>>
>>>>> On Fri, Nov 27, 2020 at 9:51 PM Fujii Masao <masao.fujii@oss.nttdata.com> wrote:
>>>>>>
>>>>>>
>>>>>>
>>>>>> On 2020/11/27 18:38, Kasahara Tatsuhito wrote:
>>>>>>> Hi,
>>>>>>>
>>>>>>> On Fri, Nov 27, 2020 at 1:43 AM Fujii Masao <masao.fujii@oss.nttdata.com> wrote:
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> On 2020/11/26 10:41, Kasahara Tatsuhito wrote:
>>>>>>>>> On Wed, Nov 25, 2020 at 8:46 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:
>>>>>>>>>>
>>>>>>>>>> On Wed, Nov 25, 2020 at 4:18 PM Kasahara Tatsuhito
>>>>>>>>>> <kasahara.tatsuhito@gmail.com> wrote:
>>>>>>>>>>>
>>>>>>>>>>> Hi,
>>>>>>>>>>>
>>>>>>>>>>> On Wed, Nov 25, 2020 at 2:17 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:
>>>>>>>>>>>>
>>>>>>>>>>>> On Fri, Sep 4, 2020 at 7:50 PM Kasahara Tatsuhito
>>>>>>>>>>>> <kasahara.tatsuhito@gmail.com> wrote:
>>>>>>>>>>>>>
>>>>>>>>>>>>> Hi,
>>>>>>>>>>>>>
>>>>>>>>>>>>> On Wed, Sep 2, 2020 at 2:10 AM Kasahara Tatsuhito
>>>>>>>>>>>>> <kasahara.tatsuhito@gmail.com> wrote:
>>>>>>>>>>>>>>> I wonder if we could have table_recheck_autovac do two probes of the stats
>>>>>>>>>>>>>>> data.  First probe the existing stats data, and if it shows the table to
>>>>>>>>>>>>>>> be already vacuumed, return immediately.  If not, *then* force a stats
>>>>>>>>>>>>>>> re-read, and check a second time.
>>>>>>>>>>>>>> Does the above mean that the second and subsequent table_recheck_autovac()
>>>>>>>>>>>>>> will be improved to first check using the previous refreshed statistics?
>>>>>>>>>>>>>> I think that certainly works.
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> If that's correct, I'll try to create a patch for the PoC
>>>>>>>>>>>>>
>>>>>>>>>>>>> I still don't know how to reproduce Jim's troubles, but I was able to reproduce
>>>>>>>>>>>>> what was probably a very similar problem.
>>>>>>>>>>>>>
>>>>>>>>>>>>> This problem seems to be more likely to occur in cases where you have
>>>>>>>>>>>>> a large number of tables,
>>>>>>>>>>>>> i.e., a large amount of stats, and many small tables need VACUUM at
>>>>>>>>>>>>> the same time.
>>>>>>>>>>>>>
>>>>>>>>>>>>> So I followed Tom's advice and created a patch for the PoC.
>>>>>>>>>>>>> This patch will enable a flag in the table_recheck_autovac function to use
>>>>>>>>>>>>> the existing stats next time if VACUUM (or ANALYZE) has already been done
>>>>>>>>>>>>> by another worker on the check after the stats have been updated.
>>>>>>>>>>>>> If the tables continue to require VACUUM after the refresh, then a refresh
>>>>>>>>>>>>> will be required instead of using the existing statistics.
>>>>>>>>>>>>>
>>>>>>>>>>>>> I did simple test with HEAD and HEAD + this PoC patch.
>>>>>>>>>>>>> The tests were conducted in two cases.
>>>>>>>>>>>>> (I changed few configurations. see attached scripts)
>>>>>>>>>>>>>
>>>>>>>>>>>>> 1. Normal VACUUM case
>>>>>>>>>>>>>       - SET autovacuum = off
>>>>>>>>>>>>>       - CREATE tables with 100 rows
>>>>>>>>>>>>>       - DELETE 90 rows for each tables
>>>>>>>>>>>>>       - SET autovacuum = on and restart PostgreSQL
>>>>>>>>>>>>>       - Measure the time it takes for all tables to be VACUUMed
>>>>>>>>>>>>>
>>>>>>>>>>>>> 2. Anti wrap round VACUUM case
>>>>>>>>>>>>>       - CREATE brank tables
>>>>>>>>>>>>>       - SELECT all of these tables (for generate stats)
>>>>>>>>>>>>>       - SET autovacuum_freeze_max_age to low values and restart PostgreSQL
>>>>>>>>>>>>>       - Consumes a lot of XIDs by using txid_curent()
>>>>>>>>>>>>>       - Measure the time it takes for all tables to be VACUUMed
>>>>>>>>>>>>>
>>>>>>>>>>>>> For each test case, the following results were obtained by changing
>>>>>>>>>>>>> autovacuum_max_workers parameters to 1, 2, 3(def) 5 and 10.
>>>>>>>>>>>>> Also changing num of tables to 1000, 5000, 10000 and 20000.
>>>>>>>>>>>>>
>>>>>>>>>>>>> Due to the poor VM environment (2 VCPU/4 GB), the results are a little unstable,
>>>>>>>>>>>>> but I think it's enough to ask for a trend.
>>>>>>>>>>>>>
>>>>>>>>>>>>> ===========================================================================
>>>>>>>>>>>>> [1.Normal VACUUM case]
>>>>>>>>>>>>>      tables:1000
>>>>>>>>>>>>>       autovacuum_max_workers 1:   (HEAD) 20 sec VS (with patch)  20 sec
>>>>>>>>>>>>>       autovacuum_max_workers 2:   (HEAD) 18 sec VS (with patch)  16 sec
>>>>>>>>>>>>>       autovacuum_max_workers 3:   (HEAD) 18 sec VS (with patch)  16 sec
>>>>>>>>>>>>>       autovacuum_max_workers 5:   (HEAD) 19 sec VS (with patch)  17 sec
>>>>>>>>>>>>>       autovacuum_max_workers 10:  (HEAD) 19 sec VS (with patch)  17 sec
>>>>>>>>>>>>>
>>>>>>>>>>>>>      tables:5000
>>>>>>>>>>>>>       autovacuum_max_workers 1:   (HEAD) 77 sec VS (with patch)  78 sec
>>>>>>>>>>>>>       autovacuum_max_workers 2:   (HEAD) 61 sec VS (with patch)  43 sec
>>>>>>>>>>>>>       autovacuum_max_workers 3:   (HEAD) 38 sec VS (with patch)  38 sec
>>>>>>>>>>>>>       autovacuum_max_workers 5:   (HEAD) 45 sec VS (with patch)  37 sec
>>>>>>>>>>>>>       autovacuum_max_workers 10:  (HEAD) 43 sec VS (with patch)  35 sec
>>>>>>>>>>>>>
>>>>>>>>>>>>>      tables:10000
>>>>>>>>>>>>>       autovacuum_max_workers 1:   (HEAD) 152 sec VS (with patch)  153 sec
>>>>>>>>>>>>>       autovacuum_max_workers 2:   (HEAD) 119 sec VS (with patch)   98 sec
>>>>>>>>>>>>>       autovacuum_max_workers 3:   (HEAD)  87 sec VS (with patch)   78 sec
>>>>>>>>>>>>>       autovacuum_max_workers 5:   (HEAD) 100 sec VS (with patch)   66 sec
>>>>>>>>>>>>>       autovacuum_max_workers 10:  (HEAD)  97 sec VS (with patch)   56 sec
>>>>>>>>>>>>>
>>>>>>>>>>>>>      tables:20000
>>>>>>>>>>>>>       autovacuum_max_workers 1:   (HEAD) 338 sec VS (with patch)  339 sec
>>>>>>>>>>>>>       autovacuum_max_workers 2:   (HEAD) 231 sec VS (with patch)  229 sec
>>>>>>>>>>>>>       autovacuum_max_workers 3:   (HEAD) 220 sec VS (with patch)  191 sec
>>>>>>>>>>>>>       autovacuum_max_workers 5:   (HEAD) 234 sec VS (with patch)  147 sec
>>>>>>>>>>>>>       autovacuum_max_workers 10:  (HEAD) 320 sec VS (with patch)  113 sec
>>>>>>>>>>>>>
>>>>>>>>>>>>> [2.Anti wrap round VACUUM case]
>>>>>>>>>>>>>      tables:1000
>>>>>>>>>>>>>       autovacuum_max_workers 1:   (HEAD) 19 sec VS (with patch) 18 sec
>>>>>>>>>>>>>       autovacuum_max_workers 2:   (HEAD) 14 sec VS (with patch) 15 sec
>>>>>>>>>>>>>       autovacuum_max_workers 3:   (HEAD) 14 sec VS (with patch) 14 sec
>>>>>>>>>>>>>       autovacuum_max_workers 5:   (HEAD) 14 sec VS (with patch) 16 sec
>>>>>>>>>>>>>       autovacuum_max_workers 10:  (HEAD) 16 sec VS (with patch) 14 sec
>>>>>>>>>>>>>
>>>>>>>>>>>>>      tables:5000
>>>>>>>>>>>>>       autovacuum_max_workers 1:   (HEAD) 69 sec VS (with patch) 69 sec
>>>>>>>>>>>>>       autovacuum_max_workers 2:   (HEAD) 66 sec VS (with patch) 47 sec
>>>>>>>>>>>>>       autovacuum_max_workers 3:   (HEAD) 59 sec VS (with patch) 37 sec
>>>>>>>>>>>>>       autovacuum_max_workers 5:   (HEAD) 39 sec VS (with patch) 28 sec
>>>>>>>>>>>>>       autovacuum_max_workers 10:  (HEAD) 39 sec VS (with patch) 29 sec
>>>>>>>>>>>>>
>>>>>>>>>>>>>      tables:10000
>>>>>>>>>>>>>       autovacuum_max_workers 1:   (HEAD) 139 sec VS (with patch) 138 sec
>>>>>>>>>>>>>       autovacuum_max_workers 2:   (HEAD) 130 sec VS (with patch)  86 sec
>>>>>>>>>>>>>       autovacuum_max_workers 3:   (HEAD) 120 sec VS (with patch)  68 sec
>>>>>>>>>>>>>       autovacuum_max_workers 5:   (HEAD)  96 sec VS (with patch)  41 sec
>>>>>>>>>>>>>       autovacuum_max_workers 10:  (HEAD)  90 sec VS (with patch)  39 sec
>>>>>>>>>>>>>
>>>>>>>>>>>>>      tables:20000
>>>>>>>>>>>>>       autovacuum_max_workers 1:   (HEAD) 313 sec VS (with patch) 331 sec
>>>>>>>>>>>>>       autovacuum_max_workers 2:   (HEAD) 209 sec VS (with patch) 201 sec
>>>>>>>>>>>>>       autovacuum_max_workers 3:   (HEAD) 227 sec VS (with patch) 141 sec
>>>>>>>>>>>>>       autovacuum_max_workers 5:   (HEAD) 236 sec VS (with patch)  88 sec
>>>>>>>>>>>>>       autovacuum_max_workers 10:  (HEAD) 309 sec VS (with patch)  74 sec
>>>>>>>>>>>>> ===========================================================================
>>>>>>>>>>>>>
>>>>>>>>>>>>> The cases without patch, the scalability of the worker has decreased
>>>>>>>>>>>>> as the number of tables has increased.
>>>>>>>>>>>>> In fact, the more workers there are, the longer it takes to complete
>>>>>>>>>>>>> VACUUM to all tables.
>>>>>>>>>>>>> The cases with patch, it shows good scalability with respect to the
>>>>>>>>>>>>> number of workers.
>>>>>>>>>>>>
>>>>>>>>>>>> It seems a good performance improvement even without the patch of
>>>>>>>>>>>> shared memory based stats collector.
>>>>>>>>
>>>>>>>> Sounds great!
>>>>>>>>
>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>> Note that perf top results showed that hash_search_with_hash_value,
>>>>>>>>>>>>> hash_seq_search and
>>>>>>>>>>>>> pgstat_read_statsfiles are dominant during VACUUM in all patterns,
>>>>>>>>>>>>> with or without the patch.
>>>>>>>>>>>>>
>>>>>>>>>>>>> Therefore, there is still a need to find ways to optimize the reading
>>>>>>>>>>>>> of large amounts of stats.
>>>>>>>>>>>>> However, this patch is effective in its own right, and since there are
>>>>>>>>>>>>> only a few parts to modify,
>>>>>>>>>>>>> I think it should be able to be applied to current (preferably
>>>>>>>>>>>>> pre-v13) PostgreSQL.
>>>>>>>>>>>>
>>>>>>>>>>>> +1
>>>>>>>>>>>>
>>>>>>>>>>>> +
>>>>>>>>>>>> +       /* We might be better to refresh stats */
>>>>>>>>>>>> +       use_existing_stats = false;
>>>>>>>>>>>>         }
>>>>>>>>>>>> +   else
>>>>>>>>>>>> +   {
>>>>>>>>>>>>
>>>>>>>>>>>> -   heap_freetuple(classTup);
>>>>>>>>>>>> +       heap_freetuple(classTup);
>>>>>>>>>>>> +       /* The relid has already vacuumed, so we might be better to
>>>>>>>>>>>> use exiting stats */
>>>>>>>>>>>> +       use_existing_stats = true;
>>>>>>>>>>>> +   }
>>>>>>>>>>>>
>>>>>>>>>>>> With that patch, the autovacuum process refreshes the stats in the
>>>>>>>>>>>> next check if it finds out that the table still needs to be vacuumed.
>>>>>>>>>>>> But I guess it's not necessarily true because the next table might be
>>>>>>>>>>>> vacuumed already. So I think we might want to always use the existing
>>>>>>>>>>>> for the first check. What do you think?
>>>>>>>>>>> Thanks for your comment.
>>>>>>>>>>>
>>>>>>>>>>> If we assume the case where some workers vacuum on large tables
>>>>>>>>>>> and a single worker vacuum on small tables, the processing
>>>>>>>>>>> performance of the single worker will be slightly lower if the
>>>>>>>>>>> existing statistics are checked every time.
>>>>>>>>>>>
>>>>>>>>>>> In fact, at first I tried to check the existing stats every time,
>>>>>>>>>>> but the performance was slightly worse in cases with a small number of workers.
>>>>>>>>
>>>>>>>> Do you have this benchmark result?
>>>>>>>>
>>>>>>>>
>>>>>>>>>>> (Checking the existing stats is lightweight , but at high frequency,
>>>>>>>>>>>      it affects processing performance.)
>>>>>>>>>>> Therefore, at after refresh statistics, determine whether autovac
>>>>>>>>>>> should use the existing statistics.
>>>>>>>>>>
>>>>>>>>>> Yeah, since the test you used uses a lot of small tables, if there are
>>>>>>>>>> a few workers, checking the existing stats is unlikely to return true
>>>>>>>>>> (no need to vacuum). So the cost of existing stats check ends up being
>>>>>>>>>> overhead. Not sure how slow always checking the existing stats was,
>>>>>>>>>> but given that the shared memory based stats collector patch could
>>>>>>>>>> improve the performance of refreshing stats, it might be better not to
>>>>>>>>>> check the existing stats frequently like the patch does. Anyway, I
>>>>>>>>>> think it’s better to evaluate the performance improvement with other
>>>>>>>>>> cases too.
>>>>>>>>> Yeah, I would like to see how much the performance changes in other cases.
>>>>>>>>> In addition, if the shared-based-stats patch is applied, we won't need to reload
>>>>>>>>> a huge stats file, so we will just have to check the stats on
>>>>>>>>> shared-mem every time.
>>>>>>>>> Perhaps the logic of table_recheck_autovac could be simpler.
>>>>>>>>>
>>>>>>>>>>> BTW, I found some typos in comments, so attache a  fixed version.
>>>>>>>>
>>>>>>>> The patch adds some duplicated codes into table_recheck_autovac().
>>>>>>>> It's better to make the common function performing them and make
>>>>>>>> table_recheck_autovac() call that common function, to simplify the code.
>>>>>>> Thanks for your comment.
>>>>>>> Hmm.. I've cut out the duplicate part.
>>>>>>> Attach the patch.
>>>>>>> Could you confirm that it fits your expecting?
>>>>>>
>>>>>> Yes, thanks for updataing the patch! Here are another review comments.
>>>>>>
>>>>>> +       shared = pgstat_fetch_stat_dbentry(InvalidOid);
>>>>>> +       dbentry = pgstat_fetch_stat_dbentry(MyDatabaseId);
>>>>>>
>>>>>> When using the existing stats, ISTM that these are not necessary and
>>>>>> we can reuse "shared" and "dbentry" obtained before. Right?
>>>>> Yeah, but unless autovac_refresh_stats() is called, these functions
>>>>> read the information from the
>>>>> local hash table without re-read the stats file, so the process is very light.
>>>>> Therefore, I think, it is better to keep the current logic to keep the
>>>>> code simple.
>>>>>
>>>>>>
>>>>>> +               /* We might be better to refresh stats */
>>>>>> +               use_existing_stats = false;
>>>>>>
>>>>>> I think that we should add more comments about why it's better to
>>>>>> refresh the stats in this case.
>>>>>>
>>>>>> +               /* The relid has already vacuumed, so we might be better to use existing stats */
>>>>>> +               use_existing_stats = true;
>>>>>>
>>>>>> I think that we should add more comments about why it's better to
>>>>>> reuse the stats in this case.
>>>>> I added  comments.
>>>>>
>>>>> Attache the patch.
>>>>>
>>>>
>>>> Thank you for updating the patch. Here are some small comments on the
>>>> latest (v4) patch.
>>>>
>>>> +    * So if the last time we checked a table that was already vacuumed after
>>>> +    * refres stats, check the current statistics before refreshing it.
>>>> +    */
>>>>
>>>> s/refres/refresh/
>> Thanks! fixed.
>> Attached the patch.
>>
>>>>
>>>> -----
>>>> +/* Counter to determine if statistics should be refreshed */
>>>> +static bool use_existing_stats = false;
>>>> +
>>>>
>>>> I think 'use_existing_stats' can be declared within table_recheck_autovac().
>>>>
>>>> -----
>>>> While testing the performance, I realized that the statistics are
>>>> reset every time vacuumed one table, leading to re-reading the stats
>>>> file even if 'use_existing_stats' is true. Please refer that vacuum()
>>>> eventually calls AtEOXact_PgStat() which calls to
>>>> pgstat_clear_snapshot().
>>>
>>> Good catch!
>>>
>>>
>>>> I believe that's why the performance of the
>>>> method of always checking the existing stats wasn’t good as expected.
>>>> So if we save the statistics somewhere and use it for rechecking, the
>>>> results of the performance benchmark will differ between these two
>>>> methods.
>> Thanks for you checks.
>> But, if a worker did vacuum(), that means this worker had determined
>> need vacuum in the
>> table_recheck_autovac(). So, use_existing_stats set to false, and next
>> time, refresh stats.
>> Therefore I think the current patch is fine, as we want to avoid
>> unnecessary refreshing of
>> statistics before the actual vacuum(), right?
> 
> Yes, you're right.
> 
> When I benchmarked the performance of the method of always checking
> existing stats I edited your patch so that it sets use_existing_stats
> = true even if the second check is false (i.g., vacuum is needed).
> And the result I got was worse than expected especially in the case of
> a few autovacuum workers. But it doesn't evaluate the performance of
> that method rightly as the stats snapshot is cleared every time
> vacuum. Given you had similar results, I guess you used a similar way
> when evaluating it, is it right? If so, it’s better to fix this issue
> and see how the performance benchmark results will differ.
> 
> For example, the results of the test case with 10000 tables and 1
> autovacuum worker I reported before was:
> 
> 10000 tables:
>     autovac_workers 1  : 158s,157s, 290s
> 
> But after fixing that issue in the third method (always checking the
> existing stats), the results are:

Could you tell me how you fixed that issue? You copied the stats to
somewhere as you suggested or skipped pgstat_clear_snapshot() as
I suggested?

Kasahara-san seems not to like the latter idea because it might
cause bad side effect. So we should use the former idea?

> 
> 10000 tables:
>     autovac_workers 1  : 157s,157s, 160s

Looks good number!

Regards,

-- 
Fujii Masao
Advanced Computing Technology Center
Research and Development Headquarters
NTT DATA CORPORATION