Where the time actually went.
A bar per step, with the wait before a pod started drawn apart from the work it did.
A slow run gives you one number.
The run says 41 minutes. Every step has its own duration on its card. Nothing on the screen tells you which of them was the 33.
The half hour went on the one waiting for a node with a GPU on it.
Every step on one axis
The run says 41 minutes. The timeline says which step was 33 of them.
- Parallel branches sit side by side. A fan-out of forty reads as one wide block, not forty numbers.
- A finished run keeps its timeline. Open one from last Tuesday and it looks the same.
Waiting is not running
A step that sat unscheduled for half an hour and then finished in ninety seconds is not a slow step. dagweave draws the wait apart from the work, and names what the wait was.
- Image pulls, node scale-up and lock waits all land on the queued side.
- Which mutex or semaphore it is queued behind, and the run that holds it.
- You find out whether to make the container faster or the cluster bigger.
Open a step in the tracing backend you already run
A step that took three minutes and ran for forty seconds spent the rest of it somewhere. If you run Jaeger or Tempo, every step and every run carries a link that opens the trace for that node.
- Name your backend once on the connection and dagweave builds the query.
- Tracing is off in a stock Argo install. When your controller is producing no spans at all, dagweave says so on the run and names the environment variable that turns it on, rather than showing a link into nothing. Whether your collector is receiving them is not something we can see from outside your cluster.
One run, or the trend behind it
A step that has been getting slower for a month looks fine in any single run. The executions view carries success rate, p50 and p95 for the workflow the timeline belongs to.
Stop guessing which step was slow.
dagweave is in early access. Leave an email and we will bring you on as we open it up.