feat: dt.truncate supports broadcasting lhs

[reswqa]

Probably not much, but this should also improve some performance as the duration parsing cache.

Closes #15743.

[MarcoGorelli]

thanks for doing this! is sqrt(n) a good size to use in general? should that be used in #15736 ?

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 81.35%. Comparing base (0c2783a) to head (8958d12).
Report is 2 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main   #15768   +/-   ##
=======================================
  Coverage   81.35%   81.35%           
=======================================
  Files        1379     1379           
  Lines      176619   176683   +64     
  Branches     2544     2542    -2     
=======================================
+ Hits       143686   143748   +62     
- Misses      32449    32453    +4     
+ Partials      484      482    -2     

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

[codspeed-hq]

CodSpeed Performance Report

Merging #15768 will not alter performance

_{Comparing reswqa:truncate_broadcast_left (8958d12) with main (0423fa3)}

Summary

✅ 22 untouched benchmarks

[reswqa]

Before #15736, I'm not aware that there would be such a performance penalty if the size is not good.

thanks for doing this! is sqrt(n) a good size to use in general? should that be used in #15736 ?

To be honest, I'm not totally sure as I didn't do the benchmark for the cache size, just the keep it the same as we use elsewhere(e.g. extract_all_many).

But for truncate, though, it shouldn't need to be very large IMO. I have a feeling that durations usually don't have too many unique values (i.e. cache hit rate might be high).

[MarcoGorelli]

thanks for fixing this and for adding the cache!

[MarcoGorelli]

I think there may have been a performance impact from this for when every isn't an expression (but also a perf improvement when it is)

e.g.:

import timeit
import numpy as np

setup = """

import pandas as pd
import polars as pl
import numpy as np

rng = np.random.default_rng(1)
s = pl.Series(rng.integers(-1_000_000, 1_000_000, size=1_000_000)).cast(pl.Datetime)
df = pl.DataFrame({'dt': s})
df = df.with_columns(every=pl.Series(['10s']*(len(s)//2) + ['100s']*(len(s)//2)))
"""

results = np.array(timeit.Timer(
    stmt="df.select(pl.col('dt').dt.truncate('10s'))",
    setup=setup,
    )
    .repeat(7, 3)
)/3
print(f"min: {min(results)}")
print(f"max: {max(results)}")
print(f"{np.mean(results)} +/- {np.std(results)/np.sqrt(len(results))}")

results = np.array(timeit.Timer(
    stmt="df.select(pl.col('dt').dt.truncate(pl.col('every')))",
    setup=setup,
    )
    .repeat(7, 3)
)/3
print(f"min: {min(results)}")
print(f"max: {max(results)}")
print(f"{np.mean(results)} +/- {np.std(results)/np.sqrt(len(results))}")
(.venv) marcogorelli@DESKTOP-U8OKFP3:~/scratch$

On 0.20.21, minimum timings are:

• 0.01931395099988246
• 0.04977888566645561

On 0.20.22, they are:

• 0.040716169999844475
• 0.050329010666852504

So, I think the separate paths need to be restored?