{
  "id": 679209,
  "title": "Question: Does Kaggle fully replicate the notebook runtime in submission reruns?",
  "url": "/competitions/vesuvius-challenge-surface-detection/discussion/679209",
  "author_name": "",
  "post_date": "2026-02-27T22:46:40.117461800Z",
  "votes": -1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hi all — since this competition is coming to an end and many of us are still around, I wanted to ask a quick technical question about Kaggle submission container configuration. Thanks in advance for reading and replying.</p>\n<p>I’m relatively new to Kaggle, so apologies if this is a naive question.</p>\n<p>When we run a notebook/kernel, we can choose a runtime like None / GPU (T4x2) / P100 / TPU v5e-8, and we see container limits like ~4 CPU cores, 32GB RAM, and ~20GB disk.</p>\n<p><strong>My Question:</strong> </p>\n<p>Is the <strong>competition submission (rerun/scoring) environment</strong> fully replicated from the runtime we selected for the notebook (same GPU/TPU choice, RAM, disk, CPU)? Or are there differences between the interactive notebook environment and the submission rerun container?</p>\n<p>Does <strong>TPU v5e-8</strong> provide real advantage for longer-running inference submissions (with 224 CPU ,  8 TPU and 384 GB RAM)?</p>\n<p>In this competition we ran into disk limits, runtime constraints, and scaling issues when ensembling multiple models, so I’m trying to understand what assumptions are safe to make about the submission container.</p>\n<p>Thank you for your help.</p>",
  "messages": [
    {
      "id": "3414904",
      "postDate": "02/27/2026 22:46:40",
      "content": "<p>Hi all — since this competition is coming to an end and many of us are still around, I wanted to ask a quick technical question about Kaggle submission container configuration. Thanks in advance for reading and replying.</p>\n<p>I’m relatively new to Kaggle, so apologies if this is a naive question.</p>\n<p>When we run a notebook/kernel, we can choose a runtime like None / GPU (T4x2) / P100 / TPU v5e-8, and we see container limits like ~4 CPU cores, 32GB RAM, and ~20GB disk.</p>\n<p><strong>My Question:</strong> </p>\n<p>Is the <strong>competition submission (rerun/scoring) environment</strong> fully replicated from the runtime we selected for the notebook (same GPU/TPU choice, RAM, disk, CPU)? Or are there differences between the interactive notebook environment and the submission rerun container?</p>\n<p>Does <strong>TPU v5e-8</strong> provide real advantage for longer-running inference submissions (with 224 CPU ,  8 TPU and 384 GB RAM)?</p>\n<p>In this competition we ran into disk limits, runtime constraints, and scaling issues when ensembling multiple models, so I’m trying to understand what assumptions are safe to make about the submission container.</p>\n<p>Thank you for your help.</p>",
      "rawMarkdown": "Hi all — since this competition is coming to an end and many of us are still around, I wanted to ask a quick technical question about Kaggle submission container configuration. Thanks in advance for reading and replying.\n\nI’m relatively new to Kaggle, so apologies if this is a naive question.\n\nWhen we run a notebook/kernel, we can choose a runtime like None / GPU (T4x2) / P100 / TPU v5e-8, and we see container limits like ~4 CPU cores, 32GB RAM, and ~20GB disk.\n\n**My Question:** \n\nIs the **competition submission (rerun/scoring) environment** fully replicated from the runtime we selected for the notebook (same GPU/TPU choice, RAM, disk, CPU)? Or are there differences between the interactive notebook environment and the submission rerun container?\n\nDoes **TPU v5e-8** provide real advantage for longer-running inference submissions (with 224 CPU ,  8 TPU and 384 GB RAM)?\n\nIn this competition we ran into disk limits, runtime constraints, and scaling issues when ensembling multiple models, so I’m trying to understand what assumptions are safe to make about the submission container.\n\nThank you for your help.",
      "votes": null
    },
    {
      "id": "3414914",
      "postDate": "02/27/2026 23:06:16",
      "content": "<p>Yes I think it does, but in this challenge, due to the run time, I assume the full private set was run and scores will be shared promptly. Good luck!</p>",
      "rawMarkdown": "Yes I think it does, but in this challenge, due to the run time, I assume the full private set was run and scores will be shared promptly. Good luck!",
      "votes": null
    },
    {
      "id": "3414920",
      "postDate": "02/27/2026 23:41:09",
      "content": "<p>Basically, when you submit a Notebook, it runs on complete test set, (Private + Public). But only the public portion is showed on the LB, and the corresponding Private scores are also generated, but it is hidden till the competition end date. So when this competition end, notebooks will not be run again, just after a second, the private scores will be shown automatically. The submissions that are in progress will be considered as Late Submissions.</p>\n<p>So it totally depends upon the kernal/runtime you are using while submitting, the private set will be run on the same runtime. Hope this answers your question.</p>\n<p>One thing to mention, that i observed is that their machines are sometimes buggy/slow. Means the same notebook on same runtime could sometime show variance in time on complete run.</p>",
      "rawMarkdown": "Basically, when you submit a Notebook, it runs on complete test set, (Private + Public). But only the public portion is showed on the LB, and the corresponding Private scores are also generated, but it is hidden till the competition end date. So when this competition end, notebooks will not be run again, just after a second, the private scores will be shown automatically. The submissions that are in progress will be considered as Late Submissions.\n\nSo it totally depends upon the kernal/runtime you are using while submitting, the private set will be run on the same runtime. Hope this answers your question.\n\nOne thing to mention, that i observed is that their machines are sometimes buggy/slow. Means the same notebook on same runtime could sometime show variance in time on complete run.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3414914,
      "author_name": "rob1080ti",
      "author_url": "",
      "post_date": "02/27/2026 23:06:16",
      "content": "<p>Yes I think it does, but in this challenge, due to the run time, I assume the full private set was run and scores will be shared promptly. Good luck!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3414920,
      "author_name": "muhammadibrahim3093",
      "author_url": "",
      "post_date": "02/27/2026 23:41:09",
      "content": "<p>Basically, when you submit a Notebook, it runs on complete test set, (Private + Public). But only the public portion is showed on the LB, and the corresponding Private scores are also generated, but it is hidden till the competition end date. So when this competition end, notebooks will not be run again, just after a second, the private scores will be shown automatically. The submissions that are in progress will be considered as Late Submissions.</p>\n<p>So it totally depends upon the kernal/runtime you are using while submitting, the private set will be run on the same runtime. Hope this answers your question.</p>\n<p>One thing to mention, that i observed is that their machines are sometimes buggy/slow. Means the same notebook on same runtime could sometime show variance in time on complete run.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3414904": "Hi all — since this competition is coming to an end and many of us are still around, I wanted to ask a quick technical question about Kaggle submission container configuration. Thanks in advance for reading and replying.\n\nI’m relatively new to Kaggle, so apologies if this is a naive question.\n\nWhen we run a notebook/kernel, we can choose a runtime like None / GPU (T4x2) / P100 / TPU v5e-8, and we see container limits like ~4 CPU cores, 32GB RAM, and ~20GB disk.\n\n**My Question:** \n\nIs the **competition submission (rerun/scoring) environment** fully replicated from the runtime we selected for the notebook (same GPU/TPU choice, RAM, disk, CPU)? Or are there differences between the interactive notebook environment and the submission rerun container?\n\nDoes **TPU v5e-8** provide real advantage for longer-running inference submissions (with 224 CPU ,  8 TPU and 384 GB RAM)?\n\nIn this competition we ran into disk limits, runtime constraints, and scaling issues when ensembling multiple models, so I’m trying to understand what assumptions are safe to make about the submission container.\n\nThank you for your help.",
    "3414914": "Yes I think it does, but in this challenge, due to the run time, I assume the full private set was run and scores will be shared promptly. Good luck!",
    "3414920": "Basically, when you submit a Notebook, it runs on complete test set, (Private + Public). But only the public portion is showed on the LB, and the corresponding Private scores are also generated, but it is hidden till the competition end date. So when this competition end, notebooks will not be run again, just after a second, the private scores will be shown automatically. The submissions that are in progress will be considered as Late Submissions.\n\nSo it totally depends upon the kernal/runtime you are using while submitting, the private set will be run on the same runtime. Hope this answers your question.\n\nOne thing to mention, that i observed is that their machines are sometimes buggy/slow. Means the same notebook on same runtime could sometime show variance in time on complete run."
  },
  "source": "meta"
}