Weird execution time when using filter() with multiprocessing

Hello everyone, I got a dataset that contains 9746395 rows, and the dataset structure is as follows:

    features: ['time', 'text', 'lang', 'sentiment_result'],
    num_rows: 9746395

The sentiment_result is ‚Äúpositive‚ÄĚ, ‚Äúnegative‚ÄĚ, or ‚Äúneural‚ÄĚ.
For the given time (yyyy-mm-dd for example), I want to find out that this date has how many ‚Äúpositive‚ÄĚ, ‚Äúnegative‚ÄĚ and ‚Äúneural‚ÄĚ.

and my code is as follows:

for date in tqdm(df["Date"].tolist()):
   tmp = sentiment_dataset.filter(lambda x: x["time"] == date,keep_in_memory =True,num_proc=10)   
   number_of_positives.append(len(tmp.filter(lambda x: x["sentiment_result"]=="positive",keep_in_memory = True,num_proc=10)))
   number_of_negatives.append(len(tmp.filter(lambda x: x["sentiment_result"] == "negative",keep_in_memory = True,num_proc=10)))
   number_of_neutral.append(len(tmp.filter(lambda x: x["sentiment_result"] == "neutral",keep_in_memory = True,num_proc=10)))

While the code takes around 2 minutes (for tmp and tmp.filter(‚Ķ) I just try with ‚Äėpositive‚Äô and remove ‚Äėneral‚Äô and ‚Äėnegative‚Äô because of execution time) to complete each iteration, the process sometimes freezes for no reason and takes around 8 minutes to finish.
As you can see in the following image.

Can anyone show me what the problem is and how I can resolve it?
Thank you for your help.