DiscoveryResearch
Cold Start Tuning for Python Agents on Cloud Run
Measuring where the six seconds actually go on a scale-to-zero agent container, and which mitigations are worth their monthly cost.
What I'm testing
Breaking down cold start latency: container pull, Python interpreter start, dependency import, and client library initialisation.
Finding so far
Client library init is the largest single slice — lazy-initialising the Vertex AI client until first use moves meaningful time off the critical path.
A Cloud Scheduler ping every ten minutes during business hours keeps one instance warm at a
fraction of min-instances=1.
Open question
Whether the perceived-latency fix (streaming a thinking indicator immediately) makes the infrastructure fix unnecessary for this traffic profile.
Stack
- Cloud Run
- Python
- Cloud Scheduler