Field note
Run Axis video at the edge, not in the data center
Where should you process Axis camera video when the interesting frames never leave the site?
Process Axis video at the Google Cloud edge, next to the camera — not in a distant data center. The frames that matter are already on-site; send events and clips upstream, not the whole stream.
The wire is the wrong bottleneck
An Axis camera is a sensor with an opinion. It sees a gate, a yard, a line, a face at a door. The useful moment is a few seconds long. Hauling the rest of the day across a WAN so a regional VM can “look at it later” is how visual data projects stall: bandwidth, cost, and latency all argue against the architecture.
Field notes at MikeCast start from that constraint. If the interesting pixels never intended to leave the site, the model should not wait for them in another zip code.
What belongs at the edge
Keep three things close to the camera:
- Ingest. RTSP or Axis VAPIX into a short buffer you control. Do not treat cloud object storage as the first stop.
- Inference. A small detector or tracker on a Google Cloud edge node, a GPU workstation, or an appliance that can fail without paging a region.
- The decision. A clip, a still, a count, an alarm. That is the artifact worth shipping.
Everything else is optional telemetry. Store it if you must; do not make the product wait on it.
What still belongs in Google Cloud
The edge is not a religion. Train and evaluate in the region. Keep identity, search, and long-term archive where IAM and lifecycle policy already work. Replay a hard case from object storage when a model regresses. The split is operational, not ideological: compute on the stream lives near the lens; memory of the stream lives in the cloud.
A working default
For a new Axis deployment, start narrower than a “video platform”:
- One camera family, one scene type, one event you can name in a sentence.
- A closed loop from frame to notification in under a few seconds on the local network.
- A cloud job that only runs when you promote a clip, retrain, or audit.
That is enough to learn whether the visual data is real. If it is, you will feel it in the field before the dashboard looks impressive.
Why this is a Field note
MikeCast field notes are the working notes from cameras, edge boxes, and Google Cloud — not a product pitch. The bison faces the weather that is actually there. On a site, that weather is packet loss, backlight, and a camera that cannot wait for a region to warm up.