[Bug] malloc issue when using obs_query

[royassis]

Describe the bug

attr('sample') == val('xxx')
python: malloc.c:2617: sysmalloc: Assertion `(old_top == initial_top (av) && old_size == 0) || ((unsigned long) (old_size) >= MINSIZE && prev_inuse (old_top) && ((unsigned long) old_end & (pagesize - 1)) == 0)' failed.
Aborted

To Reproduce

import tiledbsoma

exp = tiledbsoma.Experiment.open("s3://xxx")
# To anndata
measurement_name = "RNA"

print(len(exp.obs)) #2343203
print(len(exp.ms["RNA"]["var"])) #22653

## Var Query
gene_names = ["xxx"]
var_query = tiledbsoma.AxisQuery(value_filter=f'var_id in {gene_names}')

## Obs Query
first_obs_name="sample"
first_obs_value="xxx"


value_filter=f"attr('{first_obs_name}') == val('{first_obs_value}')"
print(value_filter)


obs_query = tiledbsoma.AxisQuery(value_filter=value_filter)

# Run query
query = exp.axis_query("RNA", obs_query=obs_query, var_query=var_query)
adata = query.to_anndata(X_name="data")

Versions (please complete the following information):

tiledbsoma.__version__              2.3.0
TileDB core version (libtiledbsoma) 2.30.0
python version                      3.13.1.final.0
OS version                          Linux 6.8.0-1050-aws

Additional context
When I remove the obs_query the error doesn't happen
query = exp.axis_query("RNA", var_query=var_query)

This is the platform_config used with tiledbsoma.io.from_anndata to create the data

max_n_obs = 1000
if len(ad.obs) > max_n_obs:
	obs_tile_extent = int(len(ad.obs) / 10)
else:
	obs_tile_extent = int(len(ad.obs))

var_tile_extent = 5

platform_config = tiledbsoma.options.TileDBCreateOptions(
	tile_order="col-major",
	cell_order="col-major",
	dims={
		"soma_dim_0": {"filters": [{"_type": "ZstdFilter", "level": 3}], "tile": obs_tile_extent},
		"soma_dim_1": {"filters": [{"_type": "ZstdFilter", "level": 3}], "tile": var_tile_extent},
	},
	attrs={
		"soma_data": {"filters": [{"_type": "ZstdFilter", "level": 3}]},
	},
)

[royassis]

Doing consolidation and vacumming on all tiledb datasets in the tiledbsoma dataset fixed mem issue.
But the runtime is still much greater than using var_query alone.

[aaronwolen]

Thanks (as always) for the detailed report, @royassis. This is definitely something we want to chase down. A few things will help us narrow it down:

1. Fragment count on the obs array. Could you run:

   import tiledb
   uri = exp.obs.uri
   with tiledb.open(uri) as A:
       print(len(A.fragments_info))

   The fact that consolidation + vacuuming made the crash disappear suggests this is fragment-related. Your platform_config sets obs_tile_extent = len(ad.obs) / 10 per ingest, so if this experiment was built by appending many small AnnData files, you may have ended up with a large number of fragments and mismatched tile extents.

2. A backtrace. If you can run under gdb:

   gdb --args python your_script.py
   (gdb) run
   (gdb) bt

   Even a partial backtrace from the abort would tell us whether this is
   in the query-condition path, the Arrow conversion, or somewhere else.

3. Does the simpler value-filter form crash the same way?

   obs_query = tiledbsoma.AxisQuery(value_filter=f"{first_obs_name} == '{first_obs_value}'")

   1. Your attr(...) == val(...) query is valid, but I want to rule out anything specific to that syntax.
   2. If you're able to share (privately is fine) a small reproducer or even just the obs array on its own, that would let us reproduce locally.

[royassis]

Hey @aaronwolen

Hope everything is good on your side.

Sadly I can't reproduce the issue after consolidating the .obs dataframe. It won't crash.
Also recreating the initial tiledbsoma object would take some time.

[aaronwolen]

One quick question that will help us with an internal repro: was this experiment created by appending multiple anndata/h5ads into a single SOMA experiment, or was it built from a single large input in one shot?

[royassis]

a single large input in one shot.