Performance regression in fillna

[TomAugspurger]

https://pandas.pydata.org/speed/pandas/index.html#replace.FillNa.time_fillna?python=3.8&Cython=0.29.21&p-inplace=False&commits=2eb353063d6bc4f2ed22e9943d26465e284b8393-68f53cd9e556698fe786aefe90466f233eb55841

2eb3530...68f53cd

• I have checked that this issue has not already been reported.

• I have confirmed this bug exists on the latest version of pandas.

• (optional) I have confirmed this bug exists on the master branch of pandas.

---

[jorisvandenbossche]

cc @jbrockmendel

[jbrockmendel]

im seeing a 19% slowdown vs 1.1.4, vs more than 2x on the link. can anyone else reproduce either of these measurements?

[jorisvandenbossche]

So the snippet from the benchmark is basically:

N = 10 ** 6
rng = pd.date_range("1/1/2000", periods=N, freq="min")
data = np.random.randn(N)
data[::2] = np.nan
ts = pd.Series(data, index=rng)

%timeit ts.fillna(0.0)

Locally I see a 2x slowdown on master vs 1.1.

Comparing the snakeviz profiles, the additional times seems to come from _putmask_simpe

[jorisvandenbossche]

I suppose the main difference is that np.putmask is actually faster as __setitem__ (with the data from example above):

In [6]: mask = np.isnan(data)

In [8]: %timeit np.putmask(data, mask, 0.0)
498 µs ± 43.8 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

In [9]: %timeit data[mask]= 0.0
2.43 ms ± 4.34 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

(now, I don't know if this difference scales with larger data)

[jbrockmendel]

I suppose the main difference is that np.putmask is actually faster as setitem (with the data from example above):

Huh, good to know! ill make a PR to update putmask_simple and see if that solves the issue

[jbrockmendel]

By updating putmask_simple I get this particular example to slightly faster than 1.1.4. WIll have to wait and see if the asv machine sees the same difference as i do locally.