search

zombocom / puma_worker_killer

Automatically restart Puma cluster workers based on max RAM available
747 stars 77 forks source link

Killing multiple workers on each iteration of reaper, if needed #104

Open jeffsaremi opened 1 year ago

jeffsaremi commented 1 year ago

Current reap() method picks only one single puma worker and terminates that to free memory. However, in many cases, this would not be sufficient to bring the total memory to under max_ram. And then the system would have to wait for the subsequent cycles for this to happen.

This is not very optimal. In this proposed change, in each cycle, enough workers will be selected and terminated until the total consumed memory by puma is below the desired threshold.

The log outputs have slightly changed to indicate this. Example:

[5447] PumaWorkerKiller: Out of memory. 10 workers consuming total: 6500.63671875 mb out of max: 1980.0 mb. Releasing 4729.95703125 mb from 8 workers.
        Sending TERM to pid 5469 consuming 591.75390625 mb.
        Sending TERM to pid 5478 consuming 591.63671875 mb.
        Sending TERM to pid 5473 consuming 591.59375 mb.
        Sending TERM to pid 5471 consuming 591.3515625 mb.
        Sending TERM to pid 5547 consuming 591.171875 mb.
        Sending TERM to pid 5545 consuming 590.98046875 mb.
        Sending TERM to pid 5549 consuming 590.88671875 mb.
        Sending TERM to pid 5623 consuming 590.58203125 mb.
schneems commented 1 year ago

Seems like a good idea. Is this original code tested? I would like to encode this expected behavior in a test.

jeffsaremi commented 1 year ago

Seems like a good idea. Is this original code tested? I would like to encode this expected behavior in a test.

Hi, thank you! I have tested this on my desktop. I added a simple grep test to one of the tests. I also debug printed the spawn log and I could see that it was killing multiple workers at a time. I wasn't able to write a more comprehensive test. Sorry I'm still new to ruby.