Incremental recompute vs full reindex

How to run

• python benchmarks/run.py --write
• Or: PIXELTABLE_HOME=$(mktemp -d) python benchmarks/scenarios/incremental_recompute.py --rows 2000 --changed 200 --out /tmp/inc.json --scale reduced

Results

Insert of 2000 rows: 0.14 s wall, 0.14 s CPU. Initial embedding index: 0.41 s wall, 0.15 s CPU.

Update of 200 rows: 0.059 s wall, 0.022 s CPU. Full column recompute: 0.45 s wall, 0.16 s CPU. Drop index: 0.012 s wall, 0.0057 s CPU. Full reindex: 0.39 s wall, 0.14 s CPU.

Rows

2000

Changed

200

Insert

0.14 s wall, 0.14 s CPU

Initial embedding index

0.41 s wall, 0.15 s CPU

Incremental update

0.059 s wall, 0.022 s CPU

Full recompute

0.45 s wall, 0.16 s CPU

Full reindex

0.39 s wall, 0.14 s CPU

Methods

Scale of this page

reduced: 2000 rows, change 200

Full scale (not run here)

1,000,000 rows, change 100,000. Documented in benchmarks/README.md.

UDF

deterministic 32-d numpy stand-in, not a hosted model

Incremental step

table.update on id < changed, which cascades to the computed column and the index

Full recompute

table.recompute_columns on the embedding column

Full reindex

drop_embedding_index then add_embedding_index

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

• The UDF is cheap on purpose, so wall times are fractions of a second. A hosted embedding model would dominate both sides.
• Each step ran once, with no warmup, so every time is one sample. scripts/release-bench repeats its runs.
• Intel(R) Xeon(R) Processor, 4 cores, 15.6 GB, Linux 6.12.94+, Python 3.12.3, pixeltable 0.7.21. Do not compare these times to the /engineering Mac numbers.
• 1,000,000 rows were not run.

Where Pixeltable fits

If you already rebuild an external index on a schedule and that is the contract, Pixeltable is the wrong tool. Use it when a changed source row should recompute only its dependents.

import pixeltable as pxt
from pixeltable.functions.huggingface import sentence_transformer

TableModel = pxt.model_base()
text_embed = sentence_transformer.using(model_id='intfloat/e5-large-v2')

class Passages(TableModel, name='passages'):
    text: pxt.String
    __indexes__ = [
        pxt.EmbeddingIndex(text, string_embed=text_embed),
    ]

Questions