bug(python): offset overflow when issing table update

[alexkohler]

LanceDB version

0.6.8

What happened?

I seem to be running into an offset overflow when issuing an update spanning my entire table:

>> import lancedb
>> import os
>> os.environ["RUST_BACKTRACE"] = 1
>> db = lancedb.connect("/data/lance_local")
>>> table = db.open_table("05-10-testing")
>>> print(len(table))
92787275
>>> table.update(where="test_train_split = 'test'", values={"test_train_split": 'TEST'})

thread 'lance_background_thread' panicked at /root/.cargo/registry/src/index.crates.io-6f17d22bba15001f/arrow-data-50.0.0/src/transform/utils.rs:42:56:
offset overflow
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/akohler/.local/lib/python3.10/site-packages/lancedb/table.py", line 1600, in update
    self._dataset_mut.update(values_sql, where)
  File "/home/akohler/.local/lib/python3.10/site-packages/lance/dataset.py", line 995, in update
    self._ds.update(updates, where)
OSError: LanceError(IO): Execution error: External error: Execution error: ExecNode(Take): thread panicked: task 1103859 panicked, /home/runner/work/lance/lance/rust/lance-datafusion/src/chunker.rs:58:46

Any guidance on how to potentially work around this/apply my updates in smaller batches? Happy to provide additional info.

Are there known steps to reproduce?

No response

[wjones127]

It's unclear where this is coming from.

I tried to reproduce this in Lance, but it worked fine. Is there any important details I might be missing? Here's my script:

import lance
import pyarrow as pa
import random
import string
import tqdm

def rand_string(n):
    return ''.join(random.choices(string.ascii_lowercase +
                             string.digits, k=n))

# Create a batch with 100MB of string data 
data = pa.table({
    "text": pa.array([rand_string(100 * 1024) for _ in range(1024)]),
})

# Write over 5GB of data
for _ in tqdm.tqdm(range(500)):
    ds = lance.write_dataset(data, "test", mode="append")

# Try running an update query
ds.update(updates={"text": "'hello'"}, where="text = '{}'".format(data['text'][0].as_py()))

[alexkohler]

Hm, my table is pretty wide (14 columns). Would that potentially come into play when walking through the pages during the update?

[wjones127]

It's possible is has something to do with that. Could you share what operations you did to write your table? That would help me figure out how to reproduce this.

[alexkohler]

This table has grown a lot over time (and hence has a lot of versions, although I've periodically cleaned those up using cleanup_old_version(...)). Also possibly relevant: the field where I was running into this was the only column I added using add_columns and then backfilled using an update call. I unfortunately can't share the table, but I'm happy to help with debugging this any way I can beyond that!