Re: Perform streaming logical transactions by background workers and parallel apply - Mailing list pgsql-hackers

On Fri, Apr 8, 2022 at 2:44 PM houzj.fnst@fujitsu.com
<houzj.fnst@fujitsu.com> wrote:
>
> On Wednesday, April 6, 2022 1:20 PM Amit Kapila <amit.kapila16@gmail.com> wrote:
>
> > In this email, I would like to discuss allowing streaming logical
> > transactions (large in-progress transactions) by background workers
> > and parallel apply in general. The goal of this work is to improve the
> > performance of the apply work in logical replication.
> >
> > Currently, for large transactions, the publisher sends the data in
> > multiple streams (changes divided into chunks depending upon
> > logical_decoding_work_mem), and then on the subscriber-side, the apply
> > worker writes the changes into temporary files and once it receives
> > the commit, it read from the file and apply the entire transaction. To
> > improve the performance of such transactions, we can instead allow
> > them to be applied via background workers. There could be multiple
> > ways to achieve this:
> >
> > Approach-1: Assign a new bgworker (if available) as soon as the xact's
> > first stream came and the main apply worker will send changes to this
> > new worker via shared memory. We keep this worker assigned till the
> > transaction commit came and also wait for the worker to finish at
> > commit. This preserves commit ordering and avoid writing to and
> > reading from file in most cases. We still need to spill if there is no
> > worker available. We also need to allow stream_stop to complete by the
> > background worker to finish it to avoid deadlocks because T-1's
> > current stream of changes can update rows in conflicting order with
> > T-2's next stream of changes.
> >
>
> Attach the POC patch for the Approach-1 of "Perform streaming logical
> transactions by background workers". The patch is still a WIP patch as
> there are serval TODO items left, including:
>
> * error handling for bgworker
> * support for SKIP the transaction in bgworker
> * handle the case when there is no more worker available
>   (might need spill the data to the temp file in this case)
> * some potential bugs
>
> The original patch is borrowed from an old thread[1] and was rebased and
> extended/cleaned by me. Comments and suggestions are welcome.
>
> [1] https://www.postgresql.org/message-id/8eda5118-2dd0-79a1-4fe9-eec7e334de17%40postgrespro.ru
>
> Here are some performance results of the patch shared by Shi Yu off-list.
>
> The performance was tested by varying
> logical_decoding_work_mem, which include two cases:
>
> 1) bulk insert.
> 2) create savepoint and rollback to savepoint.
>
> I used synchronous logical replication in the test, compared SQL execution
> times before and after applying the patch.
>
> The results are as follows. The bar charts and the details of the test are
> Attached as well.
>
> RESULT - bulk insert (5kk)
> ----------------------------------
> logical_decoding_work_mem   64kB    128kB   256kB   512kB   1MB     2MB     4MB     8MB     16MB    32MB    64MB
> HEAD                        51.673  51.199  51.166  50.259  52.898  50.651  51.156  51.210  50.678  51.256  51.138
> patched                     36.198  35.123  34.223  29.198  28.712  29.090  29.709  29.408  34.367  34.716  35.439
>
> RESULT - rollback to savepoint (600k)
> ----------------------------------
> logical_decoding_work_mem   64kB    128kB   256kB   512kB   1MB     2MB     4MB     8MB     16MB    32MB    64MB
> HEAD                        31.101  31.087  30.931  31.015  30.920  31.109  30.863  31.008  30.875  30.775  29.903
> patched                     28.115  28.487  27.804  28.175  27.734  29.047  28.279  27.909  28.277  27.345  28.375
>
>
> Summary:
> 1) bulk insert
>
> For different logical_decoding_work_mem size, it takes about 30% ~ 45% less
> time, which looks good to me. After applying this patch, it seems that the
> performance is better when logical_decoding_work_mem is between 512kB and 8MB.
>
> 2) rollback to savepoint
>
> There is an improvement of about 5% ~ 10% after applying this patch.
>
> In this case, the patch spend less time handling the part that is not
> rolled back, because it saves the time writing the changes into a temporary file
> and reading the file. And for the part that is rolled back, it would spend more
> time than HEAD, because it takes more time to write to filesystem and rollback
> than writing a temporary file and truncating the file. Overall, the results looks
> good.

One comment on the design:
We should have a strategy to release the workers which have completed
applying the transactions, else even though there are some idle
workers for one of the subscriptions, it cannot be used by other
subscriptions.
Like in the following case:
Let's say max_logical_replication_workers is set to 10, if
subscription sub_1 uses all the 10 workers to apply the transactions
and all the 10 workers have finished applying the transactions and
then subscription sub_2 requests some workers for applying
transactions, subscription  sub_2 will not get any workers.
Maybe if the workers have completed applying the transactions,
subscription sub_2 should be able to get these workers in this case.

Regards,
Vignesh