`nlp.pipe(..., n_process>1)` won't return if wrapped by `tqdm()` and `zip()`

[ivyleavedtoadflax]

trafficstars

This is a pretty specific set of circumstances that I discovered today, but on the off chance that it is useful to someone, here it is.

If you include the output of nlp.pipe(...,n_process>1) in a zip() within tqdm() it will hang interminably. See below

How to reproduce the behaviour

#!/usr/bin/env python

import pandas as pd
import spacy
from tqdm import tqdm

data = [
    {"text": "I just wanna tell you how I'm feeling", "id": 0},
    {"text": "Gotta make you understand", "id": 1},
    {"text": "Never gonna give you up", "id": 2},
    {"text": "Never gonna let you down", "id": 3},
    {"text": "Never gonna run around and desert you", "id": 4},
    {"text": "Never gonna make you cry", "id": 5},
    {"text": "Never gonna say goodbye", "id": 6},
    {"text": "Never gonna tell a lie and hurt you", "id": 7},
]


df = pd.DataFrame(data)

nlp = spacy.load("en_core_web_md")

# Works with a single process

for id, doc in tqdm(
    zip(
       df["id"],
        nlp.pipe(
            df["text"],
            n_process=1,
        ),
    )
):

    print(id, doc.text)


# Works with no zip and multiple processes

for doc in tqdm(
        nlp.pipe(
            df["text"],
            n_process=2,
        ),
):

    print(doc.text)

# Hangs with multiple processes and zip

for id, doc in tqdm(
    zip(
       df["id"],
        nlp.pipe(
            df["text"],
            n_process=2,
        ),
    )
):

    print(id, doc.text)

Output:

$python script.py
0it [00:00, ?it/s]0 I just wanna tell you how I'm feeling
1 Gotta make you understand
2 Never gonna give you up
3 Never gonna let you down
4 Never gonna run around and desert you
5 Never gonna make you cry
6 Never gonna say goodbye
7 Never gonna tell a lie and hurt you
8it [00:00, 977.21it/s]
0it [00:00, ?it/s]I just wanna tell you how I'm feeling
Gotta make you understand
Never gonna give you up
Never gonna let you down
Never gonna run around and desert you
Never gonna make you cry
Never gonna say goodbye
Never gonna tell a lie and hurt you
8it [00:00, 345.82it/s]
0it [00:00, ?it/s]0 I just wanna tell you how I'm feeling
1 Gotta make you understand
2 Never gonna give you up
3 Never gonna let you down
4 Never gonna run around and desert you
5 Never gonna make you cry
6 Never gonna say goodbye
7 Never gonna tell a lie and hurt you
8it [00:00, 342.59it/s]

Your Environment

Info about spaCy

• spaCy version: 3.0.1
• Platform: Linux-5.8.0-50-generic-x86_64-with-glibc2.29
• Python version: 3.8.5
• Pipelines: en_ner_bc5cdr_md (0.4.0), en_core_web_md (3.0.0), en_ner_craft_md (0.4.0), en_ner_bionlp13cg_md (0.4.0), en_ner_jnlpba_md (0.4.0)

[adrianeboyd]

Thanks for the report, I can reproduce the behavior where it hangs.

As a workaround, I think it works if you wrap tqdm around the texts rather than on zip:

for i, doc in zip(ids, nlp.pipe(tqdm(texts), n_process=2)):
   print(doc)

[ivyleavedtoadflax]

ah nice, thanks @adrianeboyd

[adrianeboyd]

As a note, I've marked this as a bug because it shouldn't hang like this, but since there's an easy workaround it's going to be pretty low priority for us to fix.

Maybe some of the changes related to error handling have caused this? I'm not sure. In any case, it's better to use tqdm on something with a length rather than a generator.

[adrianeboyd]

As a note, we've seen that tqdm can run into deadlocks when errors are raised during the loop. With python 3.12 you can also see the new related deprecation warning related to fork and threading: https://discuss.python.org/t/concerns-regarding-deprecation-of-fork-with-alive-threads/33555

The spacy test suite would hang on all OSes with python 3.12 prior to 467c82439. (This commit is just a workaround for the test suite / common use cases. It doesn't fix the underlying issue with deadlocks and tqdm.)