POC: postgres_fdw insert batching

>
> Hi,
>
> One of the issues I'm fairly regularly reminded by users/customers is
> that inserting into tables sharded using FDWs are rather slow. We do
> even get it reported on pgsql-bugs from time to time [1].
>
> Some of the slowness / overhead is expected, doe to the latency between
> machines in the sharded setup. Even just 1ms latency will make it way
> more expensive than a single instance.
>
> But let's do a simple experiment, comparing a hash-partitioned regular
> partitions, and one with FDW partitions in the same instance. Scripts to
> run this are attached. The duration of inserting 1M rows to this table
> (average of 10 runs on my laptop) looks like this:
>
>    regular: 2872 ms
>    FDW:     64454 ms
>
> Yep, it's ~20x slower. On setup with ping latency well below 0.05ms.
> Imagine how would it look on sharded setups with 0.1ms or 1ms latency,
> which is probably where most single-DC clusters are :-(
>
> Now, the primary reason why the performance degrades like this is that
> while FDW has batching for SELECT queries (i.e. we read larger chunks of
> data from the cursors), we don't have that for INSERTs (or other DML).
> Every time you insert a row, it has to go all the way down into the
> partition synchronously.
>
> For some use cases this may be reduced by having many independent
> connnections from different users, so the per-user latency is higher but
> acceptable. But if you need to import larger amounts of data (say, a CSV
> file for analytics, ...) this may not work.
>
> Some time ago I wrote an ugly PoC adding batching, just to see how far
> would it get us, and it seems quite promising - results for he same
> INSERT benchmarks look like this:
>
>     FDW batching: 4584 ms
>
> So, rather nice improvement, I'd say ...

>
> Before I spend more time hacking on this, I have a couple open questions
> about the design, restrictions etc.
>
>
> 1) Extend the FDW API?
>
> In the patch, the batching is simply "injected" into the existing insert
> API method, i.e. ExecForeignInsert et al. I wonder if it'd be better to
> extend the API with a "batched" version of the method, so that we can
> easily determine whether the FDW supports batching or not - it would
> require changes in the callers, though. OTOH it might be useful for
> COPY, where we could do something similar to multi_insert (COPY already
> benefits from this patch, but it does not use the batching built-into
> COPY).

>
>
> 4) Expected consistency?
>
> I'm not entirely sure what are the consistency expectations for FDWs.
> Currently the FDW nodes pointing to the same server share a connection,
> so the inserted rows might be visible to other nodes. But if we only
> stash the rows in a local buffer for a while, that's no longer true. So
> maybe this breaks the consistency expectations?
>
> But maybe that's OK - I'm not sure how the prepared statements/cursors
> affect this. I can imagine restricting the batching only to plans where
> this is not an issue (single FDW node or something), but it seems rather
> fragile and undesirable.

> > 3) What about the other DML operations (DELETE/UPDATE)?
> >
> > The other DML operations could probably benefit from the batching too.
> > INSERT was good enough for a PoC, but having batching only for INSERT
> > seems somewhat asmymetric. DELETE/UPDATE seem more complicated because
> > of quals, but likely doable.
>
> Bulk INSERTs are more common in a sharded environment because of data
> load in say OLAP systems. Bulk update/delete are rare, although not
> that rare. So if an approach just supports bulk insert and not bulk
> UPDATE/DELETE that will address a large number of usecases IMO. But if
> we can make everything work together that would be good as well.

> On Tue, 30 Jun 2020 at 08:47, Etsuro Fujita <[hidden email]> wrote:
>> On Mon, Jun 29, 2020 at 7:52 PM Ashutosh Bapat
>> <[hidden email]> wrote:
>> > On Sun, Jun 28, 2020 at 8:40 PM Tomas Vondra
>> > <[hidden email]> wrote:
>>
>> > > 3) What about the other DML operations (DELETE/UPDATE)?
>> > >
>> > > The other DML operations could probably benefit from the batching too.
>> > > INSERT was good enough for a PoC, but having batching only for INSERT
>> > > seems somewhat asmymetric. DELETE/UPDATE seem more complicated because
>> > > of quals, but likely doable.
>> >
>> > Bulk INSERTs are more common in a sharded environment because of data
>> > load in say OLAP systems. Bulk update/delete are rare, although not
>> > that rare. So if an approach just supports bulk insert and not bulk
>> > UPDATE/DELETE that will address a large number of usecases IMO. But if
>> > we can make everything work together that would be good as well.
>>
>> In most cases, I think the entire UPDATE/DELETE operations would be
>> pushed down to the remote side by DirectModify.  So, I'm not sure we
>> really need the bulk UPDATE/DELETE.
> That may not be true for a partitioned table whose partitions are foreign tables. Esp. given the work that Amit Langote is doing [1]. It really depends on the ability of postgres_fdw to detect that the DML modifying each of the partitions can be pushed down. That may not come easily.

> On Tue, Jun 30, 2020 at 1:22 PM Ashutosh Bapat
> <[hidden email]> wrote:
> > On Tue, 30 Jun 2020 at 08:47, Etsuro Fujita <[hidden email]> wrote:
> >> On Mon, Jun 29, 2020 at 7:52 PM Ashutosh Bapat
> >> <[hidden email]> wrote:
> >> > On Sun, Jun 28, 2020 at 8:40 PM Tomas Vondra
> >> > <[hidden email]> wrote:
> >>
> >> > > 3) What about the other DML operations (DELETE/UPDATE)?
> >> > >
> >> > > The other DML operations could probably benefit from the batching too.
> >> > > INSERT was good enough for a PoC, but having batching only for INSERT
> >> > > seems somewhat asmymetric. DELETE/UPDATE seem more complicated because
> >> > > of quals, but likely doable.
> >> >
> >> > Bulk INSERTs are more common in a sharded environment because of data
> >> > load in say OLAP systems. Bulk update/delete are rare, although not
> >> > that rare. So if an approach just supports bulk insert and not bulk
> >> > UPDATE/DELETE that will address a large number of usecases IMO. But if
> >> > we can make everything work together that would be good as well.
> >>
> >> In most cases, I think the entire UPDATE/DELETE operations would be
> >> pushed down to the remote side by DirectModify.  So, I'm not sure we
> >> really need the bulk UPDATE/DELETE.
> > That may not be true for a partitioned table whose partitions are foreign tables. Esp. given the work that Amit Langote is doing [1]. It really depends on the ability of postgres_fdw to detect that the DML modifying each of the partitions can be pushed down. That may not come easily.
>
> While it's true that how to accommodate the DirectModify API in the
> new inherited update/delete planning approach is an open question on
> that thread, I would eventually like to find an answer to that.  That
> is, that work shouldn't result in losing the foreign partition's
> ability to use DirectModify API to optimize updates/deletes.

>On Sun, Jun 28, 2020 at 8:40 PM Tomas Vondra
><[hidden email]> wrote:
>
>>
>>     FDW batching: 4584 ms
>>
>> So, rather nice improvement, I'd say ...
>
>Very nice.
>
>>
>> Before I spend more time hacking on this, I have a couple open questions
>> about the design, restrictions etc.
>>
>>
>> 1) Extend the FDW API?
>>
>> In the patch, the batching is simply "injected" into the existing insert
>> API method, i.e. ExecForeignInsert et al. I wonder if it'd be better to
>> extend the API with a "batched" version of the method, so that we can
>> easily determine whether the FDW supports batching or not - it would
>> require changes in the callers, though. OTOH it might be useful for
>> COPY, where we could do something similar to multi_insert (COPY already
>> benefits from this patch, but it does not use the batching built-into
>> COPY).
>
>Amit Langote has pointed out a related patch being discussed on hackers at [1].
>
>That patch introduces a new API. But if we can do it without
>introducing a new API that will be good. FDWs which can support
>batching can just modify their code and don't have to implement and
>manage a new API. We already have a handful of those APIs.
>

>>
>> 2) What about the insert results?
>>
>> I'm not sure what to do about "result" status for the inserted rows. We
>> only really "stash" the rows into a buffer, so we don't know if it will
>> succeed or not. The patch simply assumes it will succeed, but that's
>> clearly wrong, and it may result in reporting a wrong number or rows.
>
>I didn't get this. We are executing an INSERT on the foreign server,
>so we get the number of rows INSERTed from that server. We should just
>add those up across batches. If there's a failure, it would abort the
>transaction, local as well as remote.
>

>>
>> The patch also disables the batching when the insert has a RETURNING
>> clause, because there's just a single slot (for the currently inserted
>> row). I suppose a "batching" method would take an array of slots.
>>
>
>It will be a rare case when a bulk load also has a RETURNING clause.
>So, we can leave with this restriction. We should try to choose a
>design which allows that restriction to be lifted in the future. But I
>doubt that restriction will be a serious one.
>
>>
>> 3) What about the other DML operations (DELETE/UPDATE)?
>>
>> The other DML operations could probably benefit from the batching too.
>> INSERT was good enough for a PoC, but having batching only for INSERT
>> seems somewhat asmymetric. DELETE/UPDATE seem more complicated because
>> of quals, but likely doable.
>
>Bulk INSERTs are more common in a sharded environment because of data
>load in say OLAP systems. Bulk update/delete are rare, although not
>that rare. So if an approach just supports bulk insert and not bulk
>UPDATE/DELETE that will address a large number of usecases IMO. But if
>we can make everything work together that would be good as well.
>
>In your patch, I see that an INSERT statement with batch is
>constructed as INSERT INTO ... VALUES (...), (...) as many values as
>the batch size. That won't work as is for UPDATE/DELETE since we can't
>pass multiple pairs of ctids and columns to be updated for each ctid
>in one statement. Maybe we could build as many UPDATE/DELETE
>statements as the size of a batch, but that would be ugly. What we
>need is a feature like a batch prepared statement in libpq similar to
>what JDBC supports
>((https://mkyong.com/jdbc/jdbc-preparedstatement-example-batch-update/).
>This will allow a single prepared statement to be executed with a
>batch of parameters, each batch corresponding to one foreign DML
>statement.
>

>>
>>
>> 3) Should we do batching for COPY insteads?
>>
>> While looking at multi_insert, I've realized it's mostly exactly what
>> the new "batching insert" API function would need to be. But it's only
>> really used in COPY, so I wonder if we should just abandon the idea of
>> batching INSERTs and do batching COPY for FDW tables.
>
>I think this won't support RETURNING as well. But if we could somehow
>use copy protocol to send the data to the foreign server and yet treat
>it as INSERT, that might work. I think we have find out which performs
>better COPY or batch INSERT.
>

>>
>> For cases that can replace INSERT with COPY this would be enough, but
>> unfortunately it does nothing for DELETE/UPDATE so I'm hesitant to do
>> this :-(
>
>Agreed, if we want to support bulk UPDATE/DELETE as well.
>
>>
>>
>> 4) Expected consistency?
>>
>> I'm not entirely sure what are the consistency expectations for FDWs.
>> Currently the FDW nodes pointing to the same server share a connection,
>> so the inserted rows might be visible to other nodes. But if we only
>> stash the rows in a local buffer for a while, that's no longer true. So
>> maybe this breaks the consistency expectations?
>>
>> But maybe that's OK - I'm not sure how the prepared statements/cursors
>> affect this. I can imagine restricting the batching only to plans where
>> this is not an issue (single FDW node or something), but it seems rather
>> fragile and undesirable.
>
>I think that area is grey. Depending upon where the cursor is
>positioned when a DML node executes a query, the data fetched from
>cursor may or may not see the effect of DML. The cursor position is
>based on the batch size so we already have problems in this area I
>think. Assuming that the DML and SELECT are independent this will
>work. So, the consistency problems exists, it will just be modulated
>by batching DML. I doubt that's related to this feature exclusively
>and should be solved independent of this feature.
>

>>
>> I was thinking about adding a GUC to enable/disable the batching at some
>> level (global, server, table, ...) but it seems like a bad match because
>> the consistency expectations likely depend on a query. There should be a
>> GUC to set the batch size, though (it's hardcoded to 100 for now).
>>
>
>Similar to fetch_size, it should foreign server, table level setting, IMO.
>
>[1] https://www.postgresql.org/message-id/flat/3d0909dc-3691-a576-208a-90986e55489f%40postgrespro.ru
>

> 3) Should we do batching for COPY insteads?
>
> While looking at multi_insert, I've realized it's mostly exactly what
> the new "batching insert" API function would need to be. But it's only
> really used in COPY, so I wonder if we should just abandon the idea of
> batching INSERTs and do batching COPY for FDW tables.
>
> For cases that can replace INSERT with COPY this would be enough, but
> unfortunately it does nothing for DELETE/UPDATE so I'm hesitant to do
> this :-(

>On 6/28/20 8:10 PM, Tomas Vondra wrote:
>>Now, the primary reason why the performance degrades like this is that
>>while FDW has batching for SELECT queries (i.e. we read larger chunks of
>>data from the cursors), we don't have that for INSERTs (or other DML).
>>Every time you insert a row, it has to go all the way down into the
>>partition synchronously.
>
>You added new fields into the PgFdwModifyState struct. Why you didn't
>reused ResultRelInfo::ri_CopyMultiInsertBuffer field and
>CopyMultiInsertBuffer machinery as storage for incoming tuples?
>

>MySQL and SQL Server has the same bulk insert syntax as Postgres, i.e., INSERT INTO table VALUES(record1), (record2), ...  Oracle doesn't have this syntax, but it can use CTE as follows:
>
>  INSERT INTO table
>  WITH t AS (
>    SELECT record1 FROM DUAL UNION ALL
>    SELECT record2 FROM DUAL UNION ALL
>    ...
>  )
>  SELECT * FROM t;
>
>And many DBMSs should have CTAS, INSERT SELECT, and INSERT SELECT record1 UNION ALL SELECT record2 ...
>

>The API would simply be:
>
>TupleTableSlot **
>ExecForeignMultiInsert(EState *estate,
>                  ResultRelInfo *rinfo,
>                  TupleTableSlot **slot,
>                  TupleTableSlot **planSlot,
>                  int numSlots);
>
>

> I measured performance in a basic non-partitioned case by modifying
> Tomas-san's scripts.  They perform an INSERT SELECT statement that
> copies one million records.  The table consists of two integer
> columns, with a primary key on one of those them.  You can run the
> attached prepare.sql to set up once.  local.sql inserts to the table
> directly, while fdw.sql inserts through a foreign table.
>
> The performance results, the average time of 5 runs,  were as follows
> on a Linux host where the average round-trip time of "ping localhost"
> was 34 us:
>
> master, local: 6.1 seconds master, fdw: 125.3 seconds patched, fdw:
> 11.1 seconds (11x improvement)
>

>
> The patch accumulates at most 100 records in ModifyTableState before
> inserting in bulk.  Also, when an input record is targeted for a
> different relation (= partition) than that for already accumulated
> records, insert the accumulated records and store the new record for
> later insert.
>
> [Issues]
>
> 1. Do we want a GUC parameter, say, max_bulk_insert_records =
> (integer), to control the number of records inserted at once? The
> range of allowed values would be between 1 and 1,000.  1 disables
> bulk insert. The possible reason of the need for this kind of
> parameter would be to limit the amount of memory used for accumulated
> records, which could be prohibitively large if each record is big.  I
> don't think this is a must, but I think we can have it.
>

> Hi,
>
> Thanks for working on this!
>
> On 11/10/20 1:45 AM, [hidden email] wrote:
>> Hello,
>>
>>
>> The attached patch implements the new bulk insert routine for
>> postgres_fdw and the executor utilizing it.  It passes make
>> check-world.
>>
>
> I haven't done any testing yet, just a quick review.
>
> I see the patch builds the "bulk" query in execute_foreign_modify. IMO
> that's something we should do earlier, when we're building the simple
> query (for 1-row inserts). I'd understand if you were concerned about
> overhead in case of 1-row inserts, trying to not plan the bulk query
> until necessary, but I'm not sure this actually helps.
>
> Or was the goal to build a query for every possible number of slots? I
> don't think that's really useful, considering it requires deallocating
> the old plan, preparing a new one, etc. IMO it should be sufficient to
> have two queries - one for 1-row inserts, one for the full batch. The
> last incomplete batch can be inserted using a loop of 1-row queries.
>
> That's what my patch was doing, but I'm not insisting on that - it just
> seems like a better approach to me. So feel free to argue why this is
> better.
>
>
>> I measured performance in a basic non-partitioned case by modifying
>> Tomas-san's scripts.  They perform an INSERT SELECT statement that
>> copies one million records.  The table consists of two integer
>> columns, with a primary key on one of those them.  You can run the
>> attached prepare.sql to set up once.  local.sql inserts to the table
>> directly, while fdw.sql inserts through a foreign table.
>>
>> The performance results, the average time of 5 runs,  were as follows
>> on a Linux host where the average round-trip time of "ping localhost"
>> was 34 us:
>>
>> master, local: 6.1 seconds master, fdw: 125.3 seconds patched, fdw:
>> 11.1 seconds (11x improvement)
>>
>
> Nice. I think we can't really get much closer to local master, so 6.1
> vs. 11.1 seconds look quite acceptable.
>
>>
>> The patch accumulates at most 100 records in ModifyTableState before
>> inserting in bulk.  Also, when an input record is targeted for a
>> different relation (= partition) than that for already accumulated
>> records, insert the accumulated records and store the new record for
>> later insert.
>>
>> [Issues]
>>
>> 1. Do we want a GUC parameter, say, max_bulk_insert_records =
>> (integer), to control the number of records inserted at once? The
>> range of allowed values would be between 1 and 1,000.  1 disables
>> bulk insert. The possible reason of the need for this kind of
>> parameter would be to limit the amount of memory used for accumulated
>> records, which could be prohibitively large if each record is big.  I
>> don't think this is a must, but I think we can have it.
>>
>
> I think it'd be good to have such GUC, even if only for testing and
> development. We should probably have a way to disable the batching,
> which the GUC could also do, I think. So +1 to have the GUC.
>
>> 2. Should we accumulate records per relation in ResultRelInfo
>> instead? That is, when inserting into a partitioned table that has
>> foreign partitions, delay insertion until a certain number of input
>> records accumulate, and then insert accumulated records per relation
>> (e.g., 50 records to relation A, 30 records to relation B, and 20
>> records to relation C.)  If we do that,
>>
>
> I think there's a chunk of text missing here? If we do that, then what?
>
> Anyway, I don't see why accumulating the records in ResultRelInfo would
> be better than what the patch does now. It seems to me like fairly
> specific to FDWs, so keeping it int FDW state seems appropriate. What
> would be the advantage of stashing it in ResultRelInfo?
>
>> * The order of insertion differs from the order of input records.  Is
>> it OK?
>>
>
> I think that's OK for most use cases, and if it's not (e.g. when there's
> something requiring the exact order of writes) then it's not possible to
> use batching. That's one of the reasons why I think we should have a GUC
> to disable the batching.
>
>> * Should the maximum count of accumulated records be applied per
>> relation or the query? When many foreign partitions belong to a
>> partitioned table, if the former is chosen, it may use much memory in
>> total.  If the latter is chosen, the records per relation could be
>> few and thus the benefit of bulk insert could be small.
>>
>
> I think it needs to be applied per relation, because that's the level at
> which we can do it easily and consistently. The whole point is to send
> data in sufficiently large chunks to minimize the communication overhead
> (latency etc.), but if you enforce it "per query" that seems hard.
>
> Imagine you're inserting data into a table with many partitions - how do
> you pick the number of rows to accumulate? The table may have 10 or 1000
> partitions, we may be inserting into all partitions or just a small
> subset, not all partitions may be foreign, etc. It seems pretty
> difficult to pick and enforce a reliable limit at the query level. But
> maybe I'm missing something and it's easier than I think?
>
> Of course, you're entirely correct enforcing this at the partition level
> may require a lot of memory. Sadly, I don't see a way around that,
> except for (a) disabling batching or (b) ordering the data to insert
> data into one partition at a time.
>

> I see the patch builds the "bulk" query in execute_foreign_modify. IMO
> that's something we should do earlier, when we're building the simple
> query (for 1-row inserts). I'd understand if you were concerned about
> overhead in case of 1-row inserts, trying to not plan the bulk query
> until necessary, but I'm not sure this actually helps.
>
> Or was the goal to build a query for every possible number of slots? I
> don't think that's really useful, considering it requires deallocating
> the old plan, preparing a new one, etc. IMO it should be sufficient to
> have two queries - one for 1-row inserts, one for the full batch. The
> last incomplete batch can be inserted using a loop of 1-row queries.
>
> That's what my patch was doing, but I'm not insisting on that - it just
> seems like a better approach to me. So feel free to argue why this is
> better.

> > * Should the maximum count of accumulated records be applied per
> > relation or the query? When many foreign partitions belong to a
> > partitioned table, if the former is chosen, it may use much memory in
> > total.  If the latter is chosen, the records per relation could be
> > few and thus the benefit of bulk insert could be small.
> >
>
> I think it needs to be applied per relation, because that's the level at
> which we can do it easily and consistently. The whole point is to send
> data in sufficiently large chunks to minimize the communication overhead
> (latency etc.), but if you enforce it "per query" that seems hard.
>
> Imagine you're inserting data into a table with many partitions - how do
> you pick the number of rows to accumulate? The table may have 10 or 1000
> partitions, we may be inserting into all partitions or just a small
> subset, not all partitions may be foreign, etc. It seems pretty
> difficult to pick and enforce a reliable limit at the query level. But
> maybe I'm missing something and it's easier than I think?
>
> Of course, you're entirely correct enforcing this at the partition level
> may require a lot of memory. Sadly, I don't see a way around that,
> except for (a) disabling batching or (b) ordering the data to insert
> data into one partition at a time.

> Two more comments regarding this:
>
> 1) If we want to be more strict about the memory consumption, we should
> probably set the limit in terms of memory, not number of rows. Currently
> the 100 rows may be 10kB or 10MB, there's no way to know. Of course,
> this is not the only place with this issue.
>
> 2) I wonder what the COPY FROM patch [1] does in this regard. I don't
> have time to check right now, but I suggest we try to do the same thing,
> if only to be consistent.
>
> [1]
> https://www.postgresql.org/message-id/flat/3d0909dc-3691-a576-208a-909
> 86e55489f%40postgrespro.ru