feat(python): Reduce scan_csv() (and friends') memory usage when using BytesIO

[itamarst]

When using scan_csv() on a BytesIO(), the hope is that it won't make a whole copy of the data if possible, e.g. by doing filtering as it reads. In fact, it currently does make a copy, making memory usage that much higher.

This PR fixes that, by using BytesIO.getvalue() instead of BytesIO.read(). I assume this will also help other scan_* functions since it's just generic infrastructure code.

To demonstrate, below is a standalone version of the test in the PR. When run with released 1.19:

Peak allocated during scan (MB) 75.24

When run with this branch:

Peak allocated during scan (MB) 0.17

Here's the script:

import io
import time
import tracemalloc
import polars as pl

tracemalloc.start()

# Create CSV that is ~70-85 MB in size:
f = io.BytesIO()
df = pl.DataFrame({"mydata": pl.int_range(0, 10_000_000, eager=True)})
df.write_csv(f)
assert 70_000_000 < f.tell() < 85_000_000
f.seek(0, 0)

current_mem = tracemalloc.get_traced_memory()[0]

# A lazy scan shouldn't make a full copy of the data:
assert (
    pl.scan_csv(f).filter(pl.col("mydata") == 9_999).collect(new_streaming=True).item()
    == 9_999
)

print(
    "Peak allocated during scan (MB)",
    (tracemalloc.get_traced_memory()[1] - current_mem) / (1024 * 1024),
)
time.sleep(1)
tracemalloc.stop()

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 79.86%. Comparing base (2e0b3a3) to head (b422136).
Report is 12 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main   #20649      +/-   ##
==========================================
+ Coverage   79.02%   79.86%   +0.84%     
==========================================
  Files        1557     1557              
  Lines      220551   220752     +201     
  Branches     2514     2527      +13     
==========================================
+ Hits       174283   176312    +2029     
+ Misses      45695    43862    -1833     
- Partials      573      578       +5     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[coastalwhite]

I was a bit worried that this would break if you changed the backing buffer afterwards. But it seems to not do that, so very nice work 😄

[itamarst]

I am fairly certain the StringIO path wasn't hit in existing usage of this function, and it's wrong because the caller expects strings to be paths... and doesn't match the function name, either.

[ritchie46]

Can we test this with a little bit less data. I assume it is noticeable with less.