Обсуждение: Pipelining INSERTs using libpq

I would like to pipeline INSERT statements.  The idea is to avoid
waiting for server round trips if the INSERT has no RETURNING clause and
runs in a transaction.  In my case, the failure of an individual INSERT
is not particularly interesting (it's a "can't happen" scenario, more or
less).  I believe this is how the X toolkit avoided network latency issues.

I wonder what's the best way to pipeline requests to the server using
the libpq API.  Historically, I have used COPY FROM STDIN instead, but
that requires (double) encoding and some client-side buffering plus
heuristics if multiple tables are filled.

It does not seem possible to use the asynchronous APIs for this purpose,
or am I missing something?

--
Florian Weimer / Red Hat Product Security Team

On Fri, Dec 21, 2012 at 4:31 AM, Florian Weimer <fweimer@redhat.com> wrote:
> I would like to pipeline INSERT statements.  The idea is to avoid waiting
> for server round trips if the INSERT has no RETURNING clause and runs in a
> transaction.  In my case, the failure of an individual INSERT is not
> particularly interesting (it's a "can't happen" scenario, more or less).  I
> believe this is how the X toolkit avoided network latency issues.
>
> I wonder what's the best way to pipeline requests to the server using the
> libpq API.  Historically, I have used COPY FROM STDIN instead, but that
> requires (double) encoding and some client-side buffering plus heuristics if
> multiple tables are filled.
>
> It does not seem possible to use the asynchronous APIs for this purpose, or
> am I missing something?

How you attack this problem depends a lot on if all your data you want
to insert is available at once or you have to wait for it from some
actor on the client side.  The purpose of asynchronous API is to allow
client side work to continue while the server is busy with the query.
So they would only help in your case if there was some kind of other
processing you needed to do to gather the data and/or prepare the
queries.  Maybe then you'd PQsend multiple insert statements with a
single call.

merlin

On 12/21/2012 03:29 PM, Merlin Moncure wrote:
> How you attack this problem depends a lot on if all your data you want
> to insert is available at once or you have to wait for it from some
> actor on the client side.  The purpose of asynchronous API is to allow
> client side work to continue while the server is busy with the query.

The client has only very little work to do until the next INSERT.

> So they would only help in your case if there was some kind of other
> processing you needed to do to gather the data and/or prepare the
> queries.  Maybe then you'd PQsend multiple insert statements with a
> single call.

I want to use parameterized queries, so I'll have to create an INSERT
statement which inserts multiple rows.  Given that it's still
stop-and-wait (even with PQsendParams), I can get through at most one
batch per RTT, so the number of rows would have to be rather large for a
cross-continental bulk load.  It's probably doable for local bulk loading.

Does the wire protocol support pipelining?  The server doesn't have to
do much to implement it. It just has to avoid discarding unexpected
bytes after the current frame and queue it for subsequent processing
instead.

(Sorry if this message arrives twice.)
--
Florian Weimer / Red Hat Product Security Team

Florian Weimer <fweimer@redhat.com> writes:
> Does the wire protocol support pipelining?

Yes, but libpq doesn't really expose the capability, because it's too
simple-minded to deal with more than one query in flight.

            regards, tom lane

On Sunday, December 23, 2012, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Florian Weimer <fweimer@redhat.com> writes:
>> Does the wire protocol support pipelining?
>
> Yes, but libpq doesn't really expose the capability, because it's too
> simple-minded to deal with more than one query in flight.
>
>                         regards, tom lane

You can do it with libpqtypes via array of records...

merlin