[ascibisz]

Problem

Graham ran into some performance issues and failures last week when running the cellpack studio. Upon further investigation, we noticed a CPU spike on our server at that time, which caused major performance for other requests at that time. After traffic died down, everything went back to normal. There are two takeaway problems that we wanted to address:

1. Our AWS ECS cluster wasn't autoscaling as expected (that's not going to be addressed in this PR, AWS configurations are not handled here, it will be addressed separately)
2. No one had hit the ECS instance in a while, so long that no packings had run on this specific instance before Graham started triggering requests (containers do restart periodically, so this case of someone being the first one to hit a fresh container is to be expected). When you're the first one to trigger a packing request for a given recipe, the server has to download the grid file from S3 for that recipe. For subsequent requests, that grid file will be cached locally. This isn't a crazy amount of work or anything, but it takes a couple seconds, so Ruge pointed out that we should just prefetch all these grid files whenever a cluster starts up, rather than just waiting until someone submits a request. That is what this PR is addressing.

Solution

When a new server instance starts up, proactively get the grid files for the example recipes on firebase (which are the recipes available in the recipe selection dropdown on cellpack studio) before we start accepting requests for new packings. This is only 4 grid files right now, so it takes like 10 seconds on startup, and it will allow for more consistent performance for users.

[github-actions]

Packing analysis report

Analysis for packing results located at cellpack/tests/outputs/test_spheres/spheresSST

Ingredient name	Encapsulating radius	Average number packed
ext_A	25	236.0

Packing image

Distance analysis

Expected minimum distance: 50.00
Actual minimum distance: 50.01

Ingredient key	Pairwise distance distribution
ext_A

[Copilot]

🟡 Changes recommended

An unavailable grid URL can abort server initialization, preventing all endpoints—including health checks—from starting and blocking prefetching of the remaining grids.

1 open issue

• Handle grid download failures without aborting app startup · New

What changed in this PR

This PR warms the server’s grid cache at startup to reduce first-request delays for Cellpack Studio’s example recipes.

Changes:

• Fetch example recipe references from Firebase and cache their grid files locally.
• Run prefetching during application initialization, before accepting packing requests.

File	Description
docker/​server.py	Adds example-grid prefetching and invokes it during server startup.

_{Review effort: Balanced}

[rugeli]

love it!