Re: Batching in executor

Hi Amit,

Thanks for the patch. I took a look over the weekend, and done a couple
experiments / benchmarks, so let me share some initial feedback (or
rather a bunch of questions I came up with).

I'll start with some general thoughts, before going into some nitpicky
comments about patches / code and perf results.

I think the general goal of the patch - reducing the per-tuple overhead
and making the executor more efficient for OLAP workloads - is very
desirable. I believe the limitations of per-row executor are one of the
reasons why attempts to implement a columnar TAM mostly failed. The
compression is nice, but it's hard to be competitive without an executor
that leverages that too. So starting with an executor, in a way that
helps even heap, seems like a good plan. So +1 to this.

While looking at the patch, I couldn't help but think about the index
prefetching stuff that I work on. It also introduces the concept of a
"batch", for passing data between an index AM and the executor. It's
interesting how different the designs are in some respects. I'm not
saying one of those designs is wrong, it's more due different goals.

For example, the index prefetching patch establishes a "shared" batch
struct, and the index AM is expected to fill it with data. After that,
the batch is managed entirely by indexam.c, with no AM calls. The only
AM-specific bit in the batch is "position", but that's used only when
advancing to the next page, etc.

This patch does things differently. IIUC, each TAM may produce it's own
"batch", which is then wrapped in a generic one. For example, heap
produces HeapBatch, and it gets wrapped in TupleBatch. But I think this
is fine. In the prefetching we chose to move all this code (walking the
batch items) from the AMs into the layer above, and make it AM agnostic.

But for the batching, we want to retain the custom format as long as
possible. Presumably, the various advantages of the TAMs are tied to the
custom/columnar storage format. Memory efficiency thanks to compression,
execution on compressed data, etc. Keeping the custom format as long as
possible is the whole point of "late materialization" (and materializing
as late as possible is one of the important details in column stores).

How far ahead have you though about these capabilities? I was wondering
about two things in particular. First, at which point do we have to
"materialize" the TupleBatch into some generic format (e.g. TupleSlots).
I get it that you want to enable passing batches between nodes, but
would those use the same "format" as the underlying scan node, or some
generic one? Second, will it be possible to execute expressions on the
custom batches (i.e. on "compressed data")? Or is it necessary to
"materialize" the batch into regular tuple slots? I realize those may
not be there "now" but maybe it'd be nice to plan for the future.

It might be worth exploring some columnar formats, and see if this
design would be a good fit. Let's say we want to process data read from
a parquet file. Would we be able to leverage the format, or would we
need to "materialize" into slots too early? Or maybe it'd be good to
look at the VCI extension [1], discussed in a nearby thread. AFAICS
that's still based on an index AM, but there were suggestions to use TAM
instead (and maybe that'd be a better choice).

The other option would be to "create batches" during execution, say by
having a new node that accumulates tuples, builds a batch and sends it
to the node above. This would help both in cases when either the lower
node does not produce batches at all, or the batches are too small (due
to filtering, aggregation, ...). Or course, it'd only win if this
increases efficiency of the upper part of the plan enough to pay for
building the batches. That can be a hard decision.

You also mentioned we could make batches larger by letting them span
multiple pages, etc. I'm not sure that's worth it - wouldn't that
substantially complicate the TAM code, which would need to pin+track
multiple buffers for each batch, etc.? Possible, but is it worth it?

I'm not sure allowing multi-page batches would actually solve the issue.
It'd help with batches at the "scan level", but presumably the batch
size in the upper nodes matters just as much. Large scan batches may
help, but hard to predict.

In the index prefetching patch we chose to keep batches 1:1 with leaf
pages, at least for now. Instead we allowed having multiple batches at
once. I'm not sure that'd be necessary for TAMs, though.

This also reminds me of LIMIT queries. The way I imagine a "batchified"
executor to work is that batches are essentially "units of work". For
example, a nested loop would grab a batch of tuples from the outer
relation, lookup inner tuples for the whole batch, and only then pass
the result batch. (I'm ignoring the cases when the batch explodes due to
duplicates.)

But what if there's a LIMIT 1 on top? Maybe it'd be enough to process
just the first tuple, and the rest of the batch is wasted work? Plenty
of (very expensive) OLAP have that, and many would likely benefit from
batching, so just disabling batching if there's LIMIT seems way too
heavy handed.

Perhaps it'd be good to gradually ramp up the batch size? Start with
small batches, and then make them larger. The index prefetching does
that too, indirectly - it reads the whole leaf page as a batch, but then
gradually ramps up the prefetch distance (well, read_stream does that).
Maybe the batching should have similar thing ...

In fact, how shall the optimizer decide whether to use batching? It's
one thing to decide whether a node can produce/consume batches, but
another thing is "should it"? With a node that "builds" a batch, this
decision would apply to even more plans, I guess.

I don't have a great answer to this, it seems like an incredibly tricky
costing issue. I'm a bit worried we might end up with something too
coarse, like "jit=on" which we know is causing problems (admittedly,
mostly due to a lot of the LLVM work being unpredictable/external). But
having some "adaptive" heuristics (like the gradual ramp up) might make
it less risky.

FWIW the current batch size limit (64 tuples) seems rather low, but it's
hard to say. It'd be good to be able to experiment with different
values, so I suggest we make this a GUC and not a hard-coded constant.

As for what to add to explain, I'd start by adding info about which
nodes are "batched" (consuming/producing batches), and some info about
the batch sizes. An average size, maybe a histogram if you want to be a
bit fancy.

I have no thoughts about the expression patches, at least not beyond
what I already wrote above. I don't know enough about that part.

[1]
https://www.postgresql.org/message-id/OS7PR01MB119648CA4E8502FE89056E56EEA7D2%40OS7PR01MB11964.jpnprd01.prod.outlook.com

Now, numbers from some microbenchmarks:

On 9/26/25 15:28, Amit Langote wrote:
>
> To evaluate the overheads and benefits, I ran microbenchmarks with
> single and multi-aggregate queries on a single table, with and without
> WHERE clauses. Tables were fully VACUUMed so visibility maps are set
> and IO costs are minimal. shared_buffers was large enough to fit the
> whole table (up to 10M rows, ~43 on each page), and all pages were
> prewarmed into cache before tests. Table schema/script is at [2].
>
> Observations from benchmarking (Detailed benchmark tables are at [3];
> below is just a high-level summary of the main patterns):
>
> * Single aggregate, no WHERE (SELECT count(*) FROM bar_N, SELECT
> sum(a) FROM bar_N): batching scan output alone improved latency by
> ~10-20%. Adding batched transition evaluation pushed gains to ~30-40%,
> especially once fmgr overhead was paid per batch instead of per row.
>
> * Single aggregate, with WHERE (WHERE a > 0 AND a < N): batching the
> qual interpreter gave a big step up, with latencies dropping by
> ~30-40% compared to batching=off.
>
> * Five aggregates, no WHERE: batching input from the child scan cut
> ~15% off runtime. Adding batched transition evaluation increased
> improvements to ~30%.
>
> * Five aggregates, with WHERE: modest gains from scan/input batching,
> but per-batch transition evaluation and batched quals brought ~20-30%
> improvement.
>
> * Across all cases, executor overheads became visible only after IO
> was minimized. Once executor cost dominated, batching consistently
> reduced CPU time, with the largest benefits coming from avoiding
> per-row fmgr calls and evaluating quals across batches.
>
> I would appreciate if others could try these patches with their own
> microbenchmarks or workloads and see if they can reproduce numbers
> similar to mine. Feedback on both the general direction and the
> details of the patches would be very helpful. In particular, patches
> 0001-0003, which add the basic batch APIs and integrate them into
> SeqScan, are intended to be the first candidates for review and
> eventual commit. Comments on the later, more experimental patches
> (aggregate input batching and expression evaluation (qual, aggregate
> transition) batching) are also welcome.
>

I tried to replicate the results, but the numbers I see are not this
good. In fact, I see a fair number of regressions (and some are not
negligible).

I'm attaching the scripts I used to build the tables / run the test. I
used the same table structure, and tried to follow the same query
pattern with 1 or 5 aggregates (I used "avg"), [0, 1, 5] where
conditions (with 100% selectivity).

I measured master vs. 0001-0003 vs. 0001-0007 (with batching on/off).
And I did that on my (relatively) new ryzen machine, and old xeon. The
behavior is quite different for the two machines, but none of them shows
such improvements. I used clang 19.0, and --with-llvm.

See the attached PDFs with a summary of the results, comparing the
results for master and the two batching branches.

The ryzen is much "smoother" - it shows almost no difference with
batching "off" (as expected). The "scan" branch (with 0001-0003) shows
an improvement of 5-10% - it's consistent, but much less than the 10-20%
you report. For the "agg" branch the benefits are much larger, but
there's also a significant regression for the largest table with 100M
rows (which is ~18GB on disk).

For xeon, the results are a bit more variable, but it affects runs both
with batching "on" and "off". The machine is just more noisy. There
seems to be a small benefit of "scan" batching (in most cases much less
than the 10-20%). The "agg" is a clear win, with up to 30-40% speedup,
and no regression similar to the ryzen.

Perhaps I did something wrong. It does not surprise me this is somewhat
CPU dependent. It's a bit sad the improvements are smaller for the newer
CPU, though.

I also tried running TPC-H. I don't have useful numbers yet, but I ran
into a segfault - see the attached backtrace. It only happens with the
batching, and only on Q22 for some reason. I initially thought it's a
bug in clang, because I saw it with clang-22 built from git, and not
with clang-14 or gcc. But since then I reproduced it with clang-19 (on
debian 13). Still could be a clang bug, of course. I've seen ~20 of
those segfaults so far, and the backtraces look exactly the same.

regards

--
Tomas Vondra