Re: PATCH: logical_work_mem and logical streaming of largein-progress transactions - Mailing list pgsql-hackers

On Wed, Aug 28, 2019 at 08:17:47PM +0300, Alexey Kondratov wrote:
>Hi Tomas,
>
>>>>>Interesting. Any idea where does the extra overhead in this particular
>>>>>case come from? It's hard to deduce that from the single flame graph,
>>>>>when I don't have anything to compare it with (i.e. the flame 
>>>>>graph for
>>>>>the "normal" case).
>>>>I guess that bottleneck is in disk operations. You can check
>>>>logical_repl_worker_new_perf.svg flame graph: disk reads (~9%) and
>>>>writes (~26%) take around 35% of CPU time in summary. To compare,
>>>>please, see attached flame graph for the following transaction:
>>>>
>>>>INSERT INTO large_text
>>>>SELECT (SELECT string_agg('x', ',')
>>>>FROM generate_series(1, 2000)) FROM generate_series(1, 1000000);
>>>>
>>>>Execution Time: 44519.816 ms
>>>>Time: 98333,642 ms (01:38,334)
>>>>
>>>>where disk IO is only ~7-8% in total. So we get very roughly the same
>>>>~x4-5 performance drop here. JFYI, I am using a machine with SSD 
>>>>for tests.
>>>>
>>>>Therefore, probably you may write changes on receiver in bigger chunks,
>>>>not each change separately.
>>>>
>>>Possibly, I/O is certainly a possible culprit, although we should be
>>>using buffered I/O and there certainly are not any fsyncs here. So I'm
>>>not sure why would it be cheaper to do the writes in batches.
>>>
>>>BTW does this mean you see the overhead on the apply side? Or are you
>>>running this on a single machine, and it's difficult to decide?
>>
>>I run this on a single machine, but walsender and worker are 
>>utilizing almost 100% of CPU per each process all the time, and at 
>>apply side I/O syscalls take about 1/3 of CPU time. Though I am 
>>still not sure, but for me this result somehow links performance 
>>drop with problems at receiver side.
>>
>>Writing in batches was just a hypothesis and to validate it I have 
>>performed test with large txn, but consisting of a smaller number of 
>>wide rows. This test does not exhibit any significant performance 
>>drop, while it was streamed too. So it seems to be valid. Anyway, I 
>>do not have other reasonable ideas beside that right now.
>
>I've checked recently this patch again and tried to elaborate it in 
>terms of performance. As a result I've implemented a new POC version 
>of the applier (attached). Almost everything in streaming logic stayed 
>intact, but apply worker is significantly different.
>
>As I wrote earlier I still claim, that spilling changes on disk at the 
>applier side adds additional overhead, but it is possible to get rid 
>of it. In my additional patch I do the following:
>
>1) Maintain a pool of additional background workers (bgworkers), that 
>are connected with main logical apply worker via shm_mq's. Each worker 
>is dedicated to the processing of specific streamed transaction.
>
>2) When we receive a streamed change for some transaction, we check 
>whether there is an existing dedicated bgworker in HTAB (xid -> 
>bgworker), or there are some in the idle list, or spawn a new one.
>
>3) We pass all changes (between STREAM START/STOP) to that bgworker 
>via shm_mq_send without intermediate waiting. However, we wait for 
>bgworker to apply the entire changes chunk at STREAM STOP, since we 
>don't want transactions reordering.
>
>4) When transaction is commited/aborted worker is being added to the 
>idle list and is waiting for reassigning message.
>
>5) I have used the same machinery with apply_dispatch in bgworkers, 
>since most of actions are practically very similar.
>
>Thus, we do not spill anything at the applier side, so transaction 
>changes are processed by bgworkers as normal backends do. In the same 
>time, changes processing is strictly serial, which prevents 
>transactions reordering and possible conflicts/anomalies. Even though 
>we trade off performance in favor of stability the result is rather 
>impressive. I have used a similar query for testing as before:
>
>EXPLAIN (ANALYZE, BUFFERS) INSERT INTO large_test (num1, num2, num3)
>    SELECT round(random()*10), random(), random()*142
>    FROM generate_series(1, 1000000) s(i);
>
>with 1kk (1000000), 3kk and 5kk rows; logical_work_mem = 64MB and 
>synchronous_standby_names = 'FIRST 1 (large_sub)'. Table schema is 
>following:
>
>CREATE TABLE large_test (
>    id serial primary key,
>    num1 bigint,
>    num2 double precision,
>    num3 double precision
>);
>
>Here are the results:
>
>-------------------------------------------------------------------
>| N | Time on master, sec | Total xact time, sec |     Ratio      |
>-------------------------------------------------------------------
>|                        On commit (master, v13)                  |
>-------------------------------------------------------------------
>| 1kk | 6.5               | 17.6                 | x2.74          |
>-------------------------------------------------------------------
>| 3kk | 21                | 55.4                 | x2.64          |
>-------------------------------------------------------------------
>| 5kk | 38.3              | 91.5                 | x2.39          |
>-------------------------------------------------------------------
>|                        Stream + spill                           |
>-------------------------------------------------------------------
>| 1kk | 5.9               | 18                   | x3             |
>-------------------------------------------------------------------
>| 3kk | 19.5              | 52.4                 | x2.7           |
>-------------------------------------------------------------------
>| 5kk | 33.3              | 86.7                 | x2.86          |
>-------------------------------------------------------------------
>|                        Stream + BGW pool                        |
>-------------------------------------------------------------------
>| 1kk | 6                 | 12                   | x2             |
>-------------------------------------------------------------------
>| 3kk | 18.5              | 30.5                 | x1.65          |
>-------------------------------------------------------------------
>| 5kk | 35.6              | 53.9                 | x1.51          |
>-------------------------------------------------------------------
>
>It seems that overhead added by synchronous replica is lower by 2-3 
>times compared with Postgres master and streaming with spilling. 
>Therefore, the original patch eliminated delay before large 
>transaction processing start by sender, while this additional patch 
>speeds up the applier side.
>
>Although the overall speed up is surely measurable, there is a room 
>for improvements yet:
>
>1) Currently bgworkers are only spawned on demand without some initial 
>pool and never stopped. Maybe we should create a small pool on 
>replication start and offload some of idle bgworkers if they exceed 
>some limit?
>
>2) Probably we can track somehow that incoming change has conflicts 
>with some of being processed xacts, so we can wait for specific 
>bgworkers only in that case?
>
>3) Since the communication between main logical apply worker and each 
>bgworker from the pool is a 'single producer --- single consumer' 
>problem, then probably it is possible to wait and set/check flags 
>without locks, but using just atomics.
>
>What do you think about this concept in general? Any concerns and 
>criticism are welcome!
>

Hi Alexey,

I'm unable to do any in-depth review of the patch over the next two weeks
or so, but I think the idea of having a pool of apply workers is sound and
can be quite beneficial for some workloads.

I don't think it matters very much whether the workers are started at the
beginning or allocated ad hoc, that's IMO a minor implementation detail.

There's one huge challenge that I however don't see mentioned in your
message or in the patch (after cursory reading) - ensuring the same commit
order, and introducing deadlocks that would not exist in single-process
apply.

Surely, we want to end up with the same commit order as on the upstream,
otherwise we might easily get different data on the subscriber. So when we
pass the large transaction to a separate process, then this process has
to wait for the other processes processing transactions that committed
first. And similarly, other processes have to wait for this process.
Depending on the commit order. I might have missed something, but I don't
see anything like that in your patch.

Essentially, this means there needs to be some sort of wait between those
apply processes, enforcing the commit order.

That however means we can easily introduce deadlocks into workloads where
the serial-apply would not have that issue - imagine multiple large
transactions, touching the same set of rows. We may ship them to different
bgworkers, and those processes may deadlock.

Of course, the deadlock detector will come around (assuming the wait is
done in a way visible to the detector) and will abort one of the
processes. But we don't know it'll abort the right one - it may easily
abort the apply process that needs to comit first, and eveyone else is
waitiing for it. Which stalls the apply forever.

regards

-- 
Tomas Vondra                  http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services