Baseten Chains have no rolling deployments and per-Chainlet autoscaling you must tune

TL;DR

Profile the full path and set autoscaling on the slowest Chainlet on its own terms, because entrypoint-level scaling will not save a starved downstream step.

Steps

1. When you build a Chain, plan promotions as atomic swaps, never as rolling releases. Declare every Chainlet through chains.depends() and annotate all run_remote parameters fully, or serialization breaks. Profile the full path and set autoscaling on the slowest Chainlet on its own terms, because entrypoint-level scaling will not save a starved downstream step. Avoid module-level global state and dynamic imports in Chainlet code, since Chainlets are distributed and replicated and globals do not behave as expected.

When to use

You are seeing this: When you build a Chain, plan promotions as atomic swaps, never as rolling releases. Use this skill when you run into "Baseten Chains have no rolling deployments and per-Chainlet autoscaling you must tune".

When not to use

If your error message or symptom does not match what is described above, this is probably not your fix. Search for your exact error text instead of forcing this one to fit.

Versions

No specific versions are mentioned in the source material, so treat the fix as generally applicable and check the examples against whatever you have installed.

Why this happens

The original report does not dig into a root cause. It documents the symptom and the fix that resolved it.

Source: Baseten org reference notes on Truss Chains document the differences that break agents coming from single-model deploys. Rolling deployments are not supported for Chains, so promotions are immediate traffic swaps with no gradual shift between versions. Chainlets cannot be instantiated directly, they must be wired with chains.depends(), and run_remote requires full type annotations or serialization silently fails. Autoscaling is per-Chainlet and independent, so a slow downstream Chainlet becomes the bottleneck unless it is tuned separately from the entrypoint.

Verified resolution: When you build a Chain, plan promotions as atomic swaps, never as rolling releases. Declare every Chainlet through chains.depends() and annotate all run_remote parameters fully, or serialization breaks. Profile the full path and set autoscaling on the slowest Chainlet on its own terms, because entrypoint-level scaling will not save a starved downstream step. Avoid module-level global state and dynamic imports in Chainlet code, since Chainlets are distributed and replicated and globals do not behave as expected.

Source URL: https://github.com/basetenlabs/baseten-skills/blob/HEAD/skills/baseten/references/truss-chains.md