Re: PATCH: Batch/pipelining support for libpq

From: Shay Rojansky <roji(at)roji(dot)org>
To: Craig Ringer <craig(at)2ndquadrant(dot)com>
Cc: PostgreSQL Hackers <pgsql-hackers(at)postgresql(dot)org>, Michael Paquier <michael(dot)paquier(at)gmail(dot)com>, Manuel Kniep <m(dot)kniep(at)web(dot)de>, "fujita(dot)etsuro(at)lab(dot)ntt(dot)co(dot)jp" <fujita(dot)etsuro(at)lab(dot)ntt(dot)co(dot)jp>
Subject: Re: PATCH: Batch/pipelining support for libpq
Date: 2016-10-14 10:09:44
Message-ID: CADT4RqBBw=XJ7geNkMdx+NmkQ4T1epPhH-yHeGpKmgNfGXWduA@mail.gmail.com
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-hackers

>
> > It has recently come to my attention that this implementation is
> problematic
> > because it forces the batch to occur within a transaction; in other
> words,
> > there's no option for a non-transactional batch.
>
> That's not strictly the case. If you explicitly BEGIN and COMMIT,
> those operations are honoured within the batch.
>

I wasn't precise in my formulation (although I think we understand each
other)... The problem I'm trying to address here is the fact that in the
"usual" batching implementation (i.e. where a single Sync message is sent
at the end of the batch), there's no support for batches which have no
transactions whatsoever (i.e. where each statement is auto-committed and
errors in earlier statements don't trigger skipping of later statements).

What's missing is implicit transactions. Usually if you send a series
> of messages they will each get their own implicit transaction. If you
> batch them, the whole lot gets an implicit transaction. This is
> because the PostgreSQL v3 protocol conflates transaction delineation
> with protocol error recovery into a single Sync message type.
>

That's right. I sent a message complaining about this conflation a while
ago:
https://www.postgresql.org/message-id/CADT4RqDdo9EcFbxwB_YO2H3BVZ0t-1qqZ%3D%2B%2BdVMnYaN6BpyUGQ%40mail.gmail.com.
There weren't any responses, although I'll add a note on the wiki on this
as a requested feature for the v4 protocol.

If the mythical v4 protocol ever happens I'd want to split Sync into
> two messages, one which is a protocol synchronisation message and
> another that is a transaction delimiter. Or give it flags or whatever.
>

Totally agree.

> In the mean time:
>
> > This can be a problem for
> > several reasons: users may want to sent off a batch of inserts, not
> caring
> > whether one of them fails (e.g. because of a unique constraint
> violation).
> > In other words, in some scenarios it may be appropriate for later batched
> > statements to be executed when an earlier batched statement raised an
> error.
> > If Sync is only sent at the very end, this isn't possible.
>
> Right, and that remains the case even with explicit transaction
> delineation, because the first ERROR causes processing of all
> subsequent messages to be skipped.
>
> The design I have in libpq allows for this by allowing the client to
> delimit batches without ending batch mode, concurrently consuming a
> stream of multiple batches. Each endbatch is a Sync. So a client that
> wants autocommit-like behavour can send a series of 1-query batches.

> I think I'll need to document this a bit better since it's more subtle
> than I properly explained.
>

Ah, I see. libpq's API is considerably more low-level than what Npgsql
needs to provide. If I understand correctly, you allow users to specify
exactly where to insert Sync messages (if at all), so that any number of
statements arbitrarily interleaved with Sync messages can be sent without
starting to read any results. If so, then the user indeed has everything
they need to control the exact transactional behavior they want (including
full auto-commit) without compromising on performance in any way (i.e. by
increasing roundtrips).

The only minor problem I can see is that PQsendEndBatch not only adds a
Sync message to the buffer, but also flushes it. This means that you may be
forcing users to needlessly flush the buffer just because they wanted to
insert a Sync. In other words, users can't send the following messages in a
single buffer/packet:
Prepare1/Bind1/Describe1/Execute1/Sync1/Prepare2/Bind2/Describe2/Execute2/Sync2
- they have to be split into different packets. Of course, this is a
relatively minor performance issue (especially when compared to the overall
performance benefits provided by batching), and providing an API
distinction between adding a Sync and flushing the buffer may
over-complicate the API. I just thought I'd mention it.

> Yes, that's what I suggest, and basically what the libpq batch
> interface does, though it expects the client to deal with the
> transaction boundaries.
>
> You will need to think hard about transaction boundaries as they
> relate to multi-statements unless nPgSQL parses out each statement
> from multi-statement strings like PgJDBC does. Otherwise a user can
> sneak in:
>
> somestatement; BEGIN; someotherstatement;
>
> or
>
> somestatement; CoMMiT; otherstatement;
>

That's a good point. I definitely don't want to depend on client-side
parsing of SQL in any way (in fact a planned feature is to allow using
Npgsql without any sort of client-side parsing/manipulation of SQL).
However, the fact that BEGIN/COMMIT can be sent in batches doesn't appear
too problematic to me.

When it's about to send a batch, Npgsql knows whether it's in an (explicit)
transaction or not (by examining the transaction indicator on the last
ReadyForQuery message it received). If it's not in an (explicit)
transaction, it automatically inserts a Sync message after every Execute.
If some statement happens to be a BEGIN, it will be executed as a normal
statement and so on. The only issue is that if an error occurs after a
sneaked-in BEGIN, all subsequent statements will fail, and all have the
Sync messages Npgsql inserted. The end result will be a series of errors
that will be raised up to the user, but this isn't fundamentally different
from the case of a simple auto-commit batch containing multiple failures
(because of unique constraint violation or whatever) - multiple errors is
something that will have to be handled in any case.

Thanks for all your comments. Npgsql's support of batches needs to be more
complicated than libpq's since it's a more high-level interface - whereas
libpq offloads some of the sending/processing complexity to the user,
Npgsql needs to take care of most of it internally (another good example is
deadlock avoidance...).

In response to

Responses

Browse pgsql-hackers by date

  From Date Subject
Next Message Mario De Frutos Dieguez 2016-10-14 10:57:43 Re: signal handling in plpython
Previous Message Heikki Linnakangas 2016-10-14 10:01:48 Re: signal handling in plpython