aws / aws/sagemaker-python-sdk

Sagemaker Endpoint vanishing without traces

未关闭
#3,252 2 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看
component: hosting type: bug
主要语言
Python
星标
2.3k
派生
1.3k
平均合并
1 天 22 小时
30 天内合并 PR
35

描述

**Describe the bug**
I'm currently using Sagemaker to host a custom ML model deployed to two accounts, homolog, and production. Both endpoints have the same entry point code and were deployed the same day. The homologation version suddenly disappeared on June 28th, leaving no traces besides the last HealthCheck ping on CloudWatch. After searching CloudTrail logs to see what could have happened, there was nothing out of the ordinary: deployed the endpoint and that was it. No delete command coming from anywhere.
I thought of it as a bug and promptly redeployed the model, on June 29th, assuming it wouldn't happen again. The issue is that on July 3rd the endpoint vanished without traces again. Same thing, no delete, no update, no renaming of anything on CloudTrail, and the only proof that it was ever on running were the CloudWatch logs and the CreateEndpoint entry on CloudTrail.

**To reproduce**
Couldn't reproduce the bug willingly. I couldn't gather any evidence that could lead me to the cause of the problem.

**Expected behavior**
For it not to vanish

**Screenshots or logs**

**System information**
A description of your system. Please provide:
- **SageMaker Python SDK version**: 2.72.1
- **Framework name (eg. PyTorch) or algorithm (eg. KMeans)**: SKLearn
- **Framework version**: 0.23
- **Python version**: 3.6
- **CPU or GPU**: ml.t2.medium
- **Custom Docker image (Y/N)**: N

**Additional context**
Some more information:

- The homologation(testing) endpoint wasn't called all that often and had big gaps between calls, they would only happen when we were testing something.
- It was deployed through an AWS Sagemaker Notebook using the Sagemaker SDK for Python
- First time the delta between a request and going offline was 8 hours, and the second time the delta was 48h.

贡献指南

打开贡献指南

调研方向

首先检查从 AWS SageMaker Notebook 进行的 SageMaker Python SDK 2.72.1 部署,然后比较 endpoint 在 6 月 28 日和 7 月 3 日消失时的 CloudTrail 和 CloudWatch 记录。完成标准是确定一个可复现的原因,或找到足够的证据来解释 endpoint 如何在没有记录 delete 或 update 的情况下消失。

由索引模型根据 Issue 内容生成。

评估

技术栈
aws, machine-learning, python
领域
cloud, machine-learning
Issue 类型
缺陷
难度
5/5
预计耗时
一周以上
活跃度
停滞
描述清晰度
需要澄清
新手友好度
20/100

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。