Graceful decommissioning on Spot/Standard pre-emptible nodes in DataProc cluster in GCP

158 Views Asked by At

We are trying to setup a DataProc cluster in GCP.

While we do, we try to have secondary nodes which are either Spot VMs or Standard pre-emptible VMs (to note: spot VMs are also pre-emptible).

When we do graceful decommissioning on the Spot / Standard pre-emptible VMs, will graceful decommissioning be in effect? (the node will not be reclaimed until the tasks running in the node are complete / the graceful decommissioning reaches its threshold time limit)

[OR]

The VMs will be immediately reclaimed back by GCP? (forceful decommissioning)

Could someone please help.

Thanks in advance.

1

There are 1 best solutions below

0
Dagang Wei On

The latter "The VMs will be immediately reclaimed back by GCP (forceful decommissioning)" is correct. Dataproc is built on top of GCE. Graceful decommissioning is a (YARN cluster level) Dataproc feature while spot/pre-emptible VMs are GCE features. Dataproc has no control over the underlying GCE behavior. You need to avoid spot/pre-emptible VMs if you need graceful decommissioning.