In my inference code, I tried to deliberately make some batches take more time(like 2~3 min). After I submitted, everything seemed fine(about 40+ min without errors). It should have run several the \"deliberate\" batches in the 40+ min. But after I went back home, the submission still resulted in Notebook Inference Server Error. So I am confused about how the time limit works.
\nCould any kaggle staff confirm the time limit?
In my inference code, I tried to deliberately make some batches take more time(like 2~3 min). After I submitted, everything seemed fine(about 40+ min without errors). It should have run several the \"deliberate\" batches in the 40+ min. But after I went back home, the submission still resulted in Notebook Inference Server Error. So I am confused about how the time limit works.
\nCould any kaggle staff confirm the time limit?
Here you can read the answer: https://www.kaggle.com/competitions/jane-street-real-time-market-data-forecasting/discussion/542022#3026394
\nIt is 60 second limit
Oh, a sad news. Thanks, mate.
", "rawMarkdown": "Oh, a sad news. Thanks, mate.", "votes": null }, { "id": "3038883", "postDate": "11/07/2024 13:49:18", "content": "Seems like it is 1 minute @trasibulo @takanashihumbert
", "rawMarkdown": "Seems like it is 1 minute @trasibulo @takanashihumbert", "votes": null }, { "id": "3039142", "postDate": "11/07/2024 17:49:41", "content": "it's 1 minute limit, however it seems that to score a notebook within 8h time limit a single call should take on average around 150 ms.
", "rawMarkdown": "it's 1 minute limit, however it seems that to score a notebook within 8h time limit a single call should take on average around 150 ms.", "votes": null }, { "id": "3039240", "postDate": "11/07/2024 20:25:38", "content": "I guess I was right here: https://www.kaggle.com/competitions/jane-street-real-time-market-data-forecasting/discussion/542550
\nHowever, the most important thing is that online learning and even extensive feature engineering sadly sounds impossible now
Another confusing thing, a submission with limit test successfully got a score after 5+ hours(a single model which should end in 1 hour), -0.0028…
", "rawMarkdown": "Another confusing thing, a submission with limit test successfully got a score after 5+ hours(a single model which should end in 1 hour), -0.0028...", "votes": null }, { "id": "3039249", "postDate": "11/07/2024 20:46:27", "content": "I'm sorry I didn't understand you
", "rawMarkdown": "I'm sorry I didn't understand you", "votes": null }, { "id": "3040990", "postDate": "11/09/2024 19:05:39", "content": "Each trading day, there are 968 (max) time_id. The test data batch is served by each time_id. If we consider 1 min limit, then we have 968 mins/ 60 mins = 16 hours (roughly). So I assume the competition organiser hopes to get the useful code which could be applied in daily high frequency trading. Under this assumption, 1min should be the real limit.
", "rawMarkdown": "Each trading day, there are 968 (max) time_id. The test data batch is served by each time_id. If we consider 1 min limit, then we have 968 mins/ 60 mins = 16 hours (roughly). So I assume the competition organiser hopes to get the useful code which could be applied in daily high frequency trading. Under this assumption, 1min should be the real limit.", "votes": null }, { "id": "3041241", "postDate": "11/10/2024 04:33:58", "content": "Isn't the test batch served by date_id? If by time_id that would make adjustment for change in the number of time_id sampled per day impossible in prediction.
", "rawMarkdown": "Isn't the test batch served by date_id? If by time_id that would make adjustment for change in the number of time_id sampled per day impossible in prediction.", "votes": null }, { "id": "3043417", "postDate": "11/12/2024 12:11:00", "content": "The total runtime of the notebook is limited to 9h so we don't have even 1 min per call, but max ~280ms, calculation is here
", "rawMarkdown": "The total runtime of the notebook is limited to 9h so we don't have even 1 min per call, but max ~280ms, calculation is [here](https://www.kaggle.com/competitions/jane-street-real-time-market-data-forecasting/discussion/545070)", "votes": null }, { "id": "3043425", "postDate": "11/12/2024 12:15:45", "content": "Lags are served once by date_id when time_id == 0 and then each time_id from the test set is served in batches for this current date_id
", "rawMarkdown": "Lags are served once by date_id when time_id == 0 and then each time_id from the test set is served in batches for this current date_id", "votes": null }, { "id": "3092899", "postDate": "01/10/2025 07:52:41", "content": "he's saying how he made a test sub, where his predict() returns > 280ms, according to the logic discussed here, this should fail, but it passed, so it's very confusing
", "rawMarkdown": "he's saying how he made a test sub, where his predict() returns > 280ms, according to the logic discussed here, this should fail, but it passed, so it's very confusing", "votes": null } ], "comments": [ { "id": 3038874, "author_name": "trasibulo", "author_url": "", "post_date": "11/07/2024 13:28:03", "content": "Here you can read the answer: https://www.kaggle.com/competitions/jane-street-real-time-market-data-forecasting/discussion/542022#3026394
\nIt is 60 second limit
Oh, a sad news. Thanks, mate.
", "votes": null, "replies": [] } ] }, { "id": 3038883, "author_name": "ravi20076", "author_url": "", "post_date": "11/07/2024 13:49:18", "content": "Seems like it is 1 minute @trasibulo @takanashihumbert
", "votes": null, "replies": [] }, { "id": 3039142, "author_name": "younesbenalia", "author_url": "", "post_date": "11/07/2024 17:49:41", "content": "it's 1 minute limit, however it seems that to score a notebook within 8h time limit a single call should take on average around 150 ms.
", "votes": null, "replies": [] }, { "id": 3039240, "author_name": "aymanallawi", "author_url": "", "post_date": "11/07/2024 20:25:38", "content": "I guess I was right here: https://www.kaggle.com/competitions/jane-street-real-time-market-data-forecasting/discussion/542550
\nHowever, the most important thing is that online learning and even extensive feature engineering sadly sounds impossible now
Another confusing thing, a submission with limit test successfully got a score after 5+ hours(a single model which should end in 1 hour), -0.0028…
", "votes": null, "replies": [ { "id": 3039249, "author_name": "aymanallawi", "author_url": "", "post_date": "11/07/2024 20:46:27", "content": "I'm sorry I didn't understand you
", "votes": null, "replies": [ { "id": 3092899, "author_name": "alexmason11", "author_url": "", "post_date": "01/10/2025 07:52:41", "content": "he's saying how he made a test sub, where his predict() returns > 280ms, according to the logic discussed here, this should fail, but it passed, so it's very confusing
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 3040990, "author_name": "gali1eo", "author_url": "", "post_date": "11/09/2024 19:05:39", "content": "Each trading day, there are 968 (max) time_id. The test data batch is served by each time_id. If we consider 1 min limit, then we have 968 mins/ 60 mins = 16 hours (roughly). So I assume the competition organiser hopes to get the useful code which could be applied in daily high frequency trading. Under this assumption, 1min should be the real limit.
", "votes": null, "replies": [ { "id": 3041241, "author_name": "casras", "author_url": "", "post_date": "11/10/2024 04:33:58", "content": "Isn't the test batch served by date_id? If by time_id that would make adjustment for change in the number of time_id sampled per day impossible in prediction.
", "votes": null, "replies": [ { "id": 3043425, "author_name": "eu1234", "author_url": "", "post_date": "11/12/2024 12:15:45", "content": "Lags are served once by date_id when time_id == 0 and then each time_id from the test set is served in batches for this current date_id
", "votes": null, "replies": [] } ] }, { "id": 3043417, "author_name": "eu1234", "author_url": "", "post_date": "11/12/2024 12:11:00", "content": "The total runtime of the notebook is limited to 9h so we don't have even 1 min per call, but max ~280ms, calculation is here
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3038870": "In my inference code, I tried to deliberately make some batches take more time(like 2~3 min). After I submitted, everything seemed fine(about 40+ min without errors). It should have run several the \"deliberate\" batches in the 40+ min. But after I went back home, the submission still resulted in Notebook Inference Server Error. So I am confused about how the time limit works.\nCould any kaggle staff confirm the time limit?", "3038874": "Here you can read the answer: https://www.kaggle.com/competitions/jane-street-real-time-market-data-forecasting/discussion/542022#3026394\nIt is 60 second limit", "3038880": "Oh, a sad news. Thanks, mate.", "3038883": "Seems like it is 1 minute @trasibulo @takanashihumbert", "3039142": "it's 1 minute limit, however it seems that to score a notebook within 8h time limit a single call should take on average around 150 ms.", "3039240": "I guess I was right here: https://www.kaggle.com/competitions/jane-street-real-time-market-data-forecasting/discussion/542550\nHowever, the most important thing is that online learning and even extensive feature engineering sadly sounds impossible now", "3039244": "Another confusing thing, a submission with limit test successfully got a score after 5+ hours(a single model which should end in 1 hour), -0.0028...", "3039249": "I'm sorry I didn't understand you", "3040990": "Each trading day, there are 968 (max) time_id. The test data batch is served by each time_id. If we consider 1 min limit, then we have 968 mins/ 60 mins = 16 hours (roughly). So I assume the competition organiser hopes to get the useful code which could be applied in daily high frequency trading. Under this assumption, 1min should be the real limit.", "3041241": "Isn't the test batch served by date_id? If by time_id that would make adjustment for change in the number of time_id sampled per day impossible in prediction.", "3043417": "The total runtime of the notebook is limited to 9h so we don't have even 1 min per call, but max ~280ms, calculation is [here](https://www.kaggle.com/competitions/jane-street-real-time-market-data-forecasting/discussion/545070)", "3043425": "Lags are served once by date_id when time_id == 0 and then each time_id from the test set is served in batches for this current date_id", "3092899": "he's saying how he made a test sub, where his predict() returns > 280ms, according to the logic discussed here, this should fail, but it passed, so it's very confusing" }, "source": "meta" }