UltraDict.apply_update() - Freezes w/ high CPU usage or AssertionError

[supermorphDotTech]

Issue Description

Hi there,

I have been troubleshooting my framework for quite some time now. As a backbone, it uses UltraDict. To make sure my software works reliably I created a stress-test with many reads, writes and of course context managers. I can not find any context manager based deadlock in my code and at this point I am suspecting UltraDict.apply_update() to be the culprit.

Although not required to discuss the issue, If you are interested you can find my repository with the rolling-update branch here: morPy

Symptom 1

Freezes with high CPU usage. The CPU utilization shows a high usage plateau when it occurs, visually understating that a context manager deadlock may be the issue.

Symptom 2

Sometimes the opportunistic reads of apply_update() seem to fail with an AssertionError.

/morPy/.venv-linux/lib/python3.10/site-packages/UltraDict/UltraDict.py", line 772, in apply_update
    assert bytes(self.buffer.buf[pos:pos+1]) == b'\xFF'
AssertionError

My Questions

A) Are there known culprits (an oversight of mine) which lead to this issue?
B) Was lock stealing (partly) meant to mitigate the freezes?
C) What can I do about the AssertionError?

[supermorphDotTech]

Proposal 1

To mitigate freezes, the lock stealing is a valid mechanism. However, it will mess with my software. I could monkey patch it, but would much rather prefer providing a pid during UltraDict construction which is exclusively eligible to steal the lock.

Proposal 2

For the AssertionError, I suggest another safety net by storing previous buffer positions to unlock repetition of apply_update() in case it fails.

My next steps

...I will see if I can manage to engineer some changes within UltraDict, but I am not sure if I can do it. In any case, you may raise concerns in respect to my proposals. You also may provide more efficient mitigations.

[supermorphDotTech]

Please pull dev into main

Currently my stress test (Linux) is running with the UltraDict dev-branch. Windows testing is pending. The test is very heavy and takes a lot of time. However, the dev-branch has long surpassed stability of the main branch and none of the errors above appeared. For now, it seems rock solid.

When will you finish current dev or at least update main with the current dev? The mutex introduced in dev worked wonders it seems.

Just please pull dev to main, since my framework apparently depends on it 🫶

[ronny-rentner]

Hey, thanks for the feedback and sorry for the delay. Will update main soon and do a new release.