Benchmarks

Computed-column UDF calls: insert, read, recompute, update

On 1000 rows a local stand-in UDF ran 1000 times on insert, 0 times when a select read the stored column, 1000 times on recompute_columns, and 100 times when 100 titles changed. No hosted LLM was called, and the page publishes call counts, not a cost. 1,000,000 rows were not run.

Updated 2026-10-28

How to run

python benchmarks/run.py --write
Or: PIXELTABLE_HOME=$(mktemp -d) python benchmarks/scenarios/udf_calls.py --rows 1000 --changed 100 --out /tmp/udf.json --scale reduced

Results

Insert: 1000 UDF calls, 0.089 s wall, 0.081 s CPU. Stored read: 0 UDF calls, 0.016 s wall, 0.014 s CPU.

Full recompute: 1000 UDF calls, 0.087 s wall, 0.079 s CPU. Update of 100 rows: 100 UDF calls, 0.022 s wall, 0.014 s CPU.

Rows: 1000
Insert UDF calls: 1000
Stored-read UDF calls: 0
Full-recompute UDF calls: 1000
Subset-update UDF calls: 100
Hosted LLM called: no

Methods

Scale of this page: reduced: 1000 rows, change 100
Full scale (not run here): 1,000,000 rows. The suite flag is --rows 1000000.
UDF: fake_llm in benchmarks/udfs.py. Writes a counter file. Returns the first 24 characters. Not an API call.
Stored read: select of the computed column after insert. It reads the stored values, so it is expected to call the UDF zero times.
Machine: Intel(R) Xeon(R) Processor, 4 cores, 15.6 GB, Linux 6.12.94+, Python 3.12.3, pixeltable 0.7.21
Recorded: 2026-10-28, command python benchmarks/run.py --write

Limitations

No hosted model ran. Call counts measure when Pixeltable runs the UDF, not GPT latency or price.
Each step ran once, with no warmup, so every time is one sample. scripts/release-bench repeats its runs.
1,000,000 rows were not run.
File I/O inside the stand-in UDF is part of the wall time. A real HTTP client would look different.

Where Pixeltable fits

If you want to fire a new HTTP call on every read, Pixeltable is the wrong tool. A computed column runs when its inputs change, and a read returns the stored value. That is what this page counted.

import pixeltable as pxt
from pixeltable.functions.openai import chat_completions

TableModel = pxt.model_base()

class Notes(TableModel, name='notes'):
    text: pxt.String
    summary = chat_completions(
        messages=[{'role': 'user', 'content': text}],
        model='gpt-4o-mini',
    ).choices[0].message.content

Questions

Does reading a computed column call the UDF again?

No. A select reads the stored value: 0 UDF calls on this run. The UDF runs on insert, on recompute_columns (1000 calls), and for the rows an update changes (100 calls).