As of now we have a following algorithm to handle node failures:
1. In case of a missing node the leader of the pipline or the scm can detected the missing heartbeats.
2. SCM will start to close the pipeline (CLOSING state) and try to close the containers with the remaining nodes in the pipeline
3. After 5 minutes the pipeline will be destroyed (CLOSED) and a new pipeline can be created from the healthy nodes (one node can be part only one pipwline in the same time).
While this algorithm can work well with a big cluster it doesn't provide very good usability on small clusters:
Given 3 nodes, in case of a service restart, if the restart takes more than 90s, the pipline will be moved to the CLOSING state. For the next 5 minutes (ozone.scm.pipeline.destroy.timeout) the container will remain in the CLOSING state. As there are no more nodes and we can't assign the same node to two different pipeline, the cluster will be unavailable for 5 minutes.
Given 90 nodes and 30 pipelines where all the pipelines are spread across 3 racks. Let's stop one rack. As all the pipelines are affected, all the pipelines will be moved to the CLOSING state. We have no free nodes, therefore we need to wait for 5 minutes to write any data to the cluster.
These problems can be solved in multiple ways:
1.) Instead of waiting 5 minutes, destroy the pipeline when all the containers are reported to be closed. (Most of the time it's enough, but some container report can be missing)
2.) Support multi-raft and open a pipeline as soon as we have enough nodes (even if the nodes already have a CLOSING pipelines).
Both the options require more work on the pipeline management side. For 0.4.0 we can adjust the following parameters to get better user experience:
First of all, we can be more optimistic and mark node to stale only after 5 mins instead of 90s. 5 mins should be enough most of the time to recover the nodes.
Second: we can decrease the time of ozone.scm.pipeline.destroy.timeout. Ideally the close command is sent by the scm to the datanode with a HB. Between two HB we have enough time to close all the containers via ratis. With the next HB, datanode can report the successful datanode. (If the containers can be closed the scm can manage the QUASI_CLOSED containers)
We need to wait 29 seconds (worst case) for the next HB, and 29+30 seconds for the confirmation. --> 66 seconds seems to be a safe choice (assuming that 6 seconds is enough to process the report about the successful closing)