Use container's RAM limit set by deploy resources limit

[iamriajul]

Describe the bug
I was having an issue of suddenly crashing meilisearch. I was using the same data sets on another server (2GiB RAM, EC2) there it had no container resource limitation, but on this new server (32 GiB Bare Metal) I deployed meilisearch with resources limitation of '8 GiB RAM Limit'.

To Reproduce

1. Deploy Meilisearch in a Docker Swarm Environment with Resources Limit (Memory Limit).

    deploy:
      resources:
        limits:
          memory: 8G # 1/4  of the available RAM Host

2. docker exec -it {containerId} sh
3. meilisearch --help
   It'll show the default meilisearch was going to use for max available, as you can see my case it defaulted to 20 GiB but I've resources limit of 8 GiB Memory.
   

Expected behavior
It should detect the maximum available memory set by the container and then automatically 2/3 of the memory from the limit instead of the host's total memory.

Workaround
As a workaround I had to redeploy meilisearch with MEILI_MAX_INDEXING_MEMORY: '6 GiB' environment configuration to consider the container's RAM Limit.

Meilisearch version:
v1.8.1 (docker)

Additional context
Architecture: x64
Docker Version: 26.1.3
Docker Mode: Swarm
Server: Bare Metal 6 cores 12 threads / 32 GiB RAM. 512 NVMe Storage.

[iamriajul]

An open issue regarding this in the sysinfo repo: GuillaumeGomez/sysinfo#207

[curquiza]

Hello @iamriajul

Thanks for the report. This is indeed an issue we know

For example when using K8s, by default Meilisearch uses

• the RAM of the node
• the CPU of the pod

The workaround is indeed to set your limit manually.

[pkruithof]

@curquiza

The workaround is indeed to set your limit manually.

I was going to report an issue, but I found this comment so I'm asking it here instead: we're trying out Meilisearch in our K8S cluster, and found that it kept on increasing memory usage over time. Up to the point it gets restarted by K8S. We've set the MEILI_MAX_INDEXING_MEMORY environment variable to 1300Mb (the memory resources are set to 1000Mi/1.95Gi), but it still shows a sawtooth graph:

The --help output shows that it's picking up the variable at least:

      --max-indexing-memory <MAX_INDEXING_MEMORY>
          Sets the maximum amount of RAM Meilisearch can use when indexing. By default, Meilisearch uses no more than two thirds of available memory
          
          [env: MEILI_MAX_INDEXING_MEMORY=1300Mb]
          [default: "10.42 GiB"]

I would have expected the line to drop — or at least stabilise — around the 1.4G mark. Is this not working correctly, or are we doing something wrong here?

[iamriajul]

@pkruithof which version of Meilisearch? There's a memory leak issue with 1.8 and up. So I'm using 1.7.6 now instead.

[pkruithof]

Ah, I didn't know that. We're using version 1.8 indeed. Do you have more information about the leak and if/when it's fixed? Maybe we'll have to downgrade for now as well.

[Kerollmops]

Hey @pkruithof 👋

@iamriajul is right there is a leak in v1.8.0 and v1.8.1. We are investigating in this Discord channel and on this GitHub issue. At this point we think we have found the root cause and are deploying a fix soon.

Sorry for the inconvenience 😞

[curquiza]

v1.8.3 with the fix has been released @iamriajul