check num_values in pg_aod_sketch_union_agg

[xinxinzhenhuai]

Context

This PR is a follow-up of #77
The test aod_sketch_test.sql failed at line 25 with Segmentation fault

 psql:test/aod_sketch_test.sql:25: server closed the connection unexpectedly
 	This probably means the server terminated abnormally
 	before or while processing the request.
 psql:test/aod_sketch_test.sql:25: error: connection to server was lost
 2025-05-23 21:14:29.609 UTC [13567] LOG:  server process (PID 13607) was terminated by signal 11: Segmentation fault
 2025-05-23 21:14:29.609 UTC [13567] DETAIL:  Failed process was running: select aod_sketch_get_estimate(aod_sketch_union(sketch, 16)) from aod_sketch_test;

The issue stems from a misuse of the aod_sketch_union function. In the following line:

select aod_sketch_get_estimate(aod_sketch_union(sketch, 16)) from aod_sketch_test;

The original intention was to set lgk to 16. However, the second argument of aod_sketch_union is actually num_values. The test table aod_sketch_test was initialized with sketches having num_values = 1 (array[1]), as shown below:

-- lgk = 16
insert into aod_sketch_test
  select aod_sketch_build(key, aod, 16)
  from (values (4, array[1]), (5, array[1]), (6, array[1]), (7, array[1]), (8, array[1])) as t(key, aod);

Passing 16 as num_values results in out-of-bounds access during the union operation, which is an undefined behavior. In our tests running on OEL7, this corrupted the array and triggered a segmentation fault. Below is the code of out-of-bound array access:

void operator()(Array& array, const Array& other) const {
  for (uint8_t i = 0; i < num_values_; ++i) {
    array[i] += other[i];
  }
}

Changes

1. Updated aod_sketch_test.sql to pass in correct lgk
2. Added num_values checks in pg_aod_sketch_union_agg

Testings

 aod_sketch_get_estimate 
-------------------------
                       8
(1 row)

 aod_sketch_get_estimate 
-------------------------
                       8
(1 row)

 aod_sketch_get_estimate 
-------------------------
                       2
(1 row)

 aod_sketch_get_estimate 
-------------------------
                       1
(1 row)

If we pass an out-of-bound num_values

 aod_sketch_get_estimate 
-------------------------
                       8
(1 row)

2025-06-06 00:25:14.352 UTC [19080] ERROR:  pg_aod_sketch_union_agg expects the same num_values in union and sketch
2025-06-06 00:25:14.352 UTC [19080] STATEMENT:  select aod_sketch_get_estimate(aod_sketch_union(sketch, 16)) from aod_sketch_test;
psql:test/aod_sketch_test.sql:25: ERROR:  pg_aod_sketch_union_agg expects the same num_values in union and sketch
 aod_sketch_get_estimate 
-------------------------
                       2
(1 row)

 aod_sketch_get_estimate 
-------------------------
                       1
(1 row)

Appendix

I have added a few log lines around the aod_union_update. Before the union operation, I could iterate over the sketch and print out its values. After the update, iterating over the sketch will produce Segmentation Fault

psql:test/aod_sketch_test.sql:25: INFO:  after update
psql:test/aod_sketch_test.sql:25: INFO:  num_values: 1
psql:test/aod_sketch_test.sql:25: INFO:  num_retained: 5
psql:test/aod_sketch_test.sql:25: INFO:  entry.first: 4611686018427387904
psql:test/aod_sketch_test.sql:25: server closed the connection unexpectedly

[AlexanderSaydakov]

something is wrong with our build:
make: /usr/bin/clang-19: No such file or directory
I am not sure where does this come from

[xinxinzhenhuai]

@AlexanderSaydakov
Hi Alexander, I have installed clang-19, then build succeed.
I feel like this is because the linux repos have updated their dependencies recently, not related to code changes.

[AlexanderSaydakov]

It is good that you found a workaround, but I am still confused as to why this clang-19 is invoked. I see that some make steps run gcc or g++. How does this happen?

[xinxinzhenhuai]

@AlexanderSaydakov
Hi Alexander. I think that's not related to the code we are adding. I believe it might be because remote repos have updated their dependencies, when installing postgresql and other packages, building tools got upgraded as well. The error got triggered while building postgres.

make: *** [/usr/lib/postgresql/16/lib/pgxs/src/makefiles/../../src/Makefile.global:1093: src/global_hooks.bc] Error 127

I have tested with an empty commit, triggered the make error as well. So the error does not come from code changes.

[AlexanderSaydakov]

I understand that it is not our code. I wonder how exactly make invokes clang-19. perhaps, pg_config suggests this here:
https://github.com/apache/datasketches-postgresql/blob/1f62c9ae6b9d8d6ca7a79032a037c22478bf82ec/Makefile#L74C1-L74C16

I am inclined to proceed with this fix for now.
Thank you.

[xinxinzhenhuai]

Thanks