If a core job fails more than the delivery limit, the leader will create a new
eval with the TriggeredBy field set to `failed-follow-up`.
Evaluations for core jobs have the leader's ACL, which is not valid on another
leader after an election. The `failed-follow-up` evals do not have ACLs, so
core job evals that fail more than the delivery limit or core job evals that
span leader elections will never succeed and will be re-enqueued forever. So
we should not retry with a `failed-follow-up`.
This fixes a bug where jobs may get "stuck" unprocessed that
dispropotionately affect periodic jobs around leadership transitions.
When registering a job, the job registration and the eval to process it
get applied to raft as two separate transactions; if the job
registration succeeds but eval application fails, the job may remain
unprocessed. Operators may detect such failure, when submitting a job
update and get a 500 error code, and they could retry; periodic jobs
failures are more likely to go unnoticed, and no further periodic
invocations will be processed until an operator force evaluation.
This fixes the issue by ensuring that the job registration and eval
application get persisted and processed atomically in the same raft log
entry.
Also, applies the same change to ensure atomicity in job deregistration.
Backward Compatibility
We must maintain compatibility in two scenarios: mixed clusters where a
leader can handle atomic updates but followers cannot, and a recent
cluster processes old log entries from legacy or mixed cluster mode.
To handle this constraints: ensure that the leader continue to emit the
Evaluation log entry until all servers have upgraded; also, when
processing raft logs, the servers honor evaluations found in both spots,
the Eval in job (de-)registration and the eval update entries.
When an updated server sees mix-mode behavior where an eval is inserted
into the raft log twice, it ignores the second instance.
I made one compromise in consistency in the mixed-mode scenario: servers
may disagree on the eval.CreateIndex value: the leader and updated
servers will report the job registration index while old servers will
report the index of the eval update log entry. This discripency doesn't
seem to be material - it's the eval.JobModifyIndex that matters.
Ensure that nomad steps down (and terminate leader goroutines) on
shutdown, when the server is the leader.
Without this change, `monitorLeadership` may handle `shutdownCh` event
and exit early before handling the raft `leaderCh` event and end up
leaking leadership goroutines.
Establishing leadership should be very fast and never make external API
calls.
This fixes a situation where there is a long backlog of Vault tokens to
be revoked on when leadership is gained. In such case, revoking the
tokens will significantly slow down leadership establishment and slow
down processing. Worse, the revocation call does not honor leadership
`stopCh` signals, so it will not stop when the leader loses leadership.
This changeset implements a periodic garbage collection of CSI volumes
with missing allocations. This can happen in a scenario where a node
update fails partially and the allocation updates are written to raft
but the evaluations to GC the volumes are dropped. This feature will
cover this edge case and ensure that upgrades from 0.11.0 and 0.11.1
get any stray claims cleaned up.
This changeset implements a periodic garbage collection of unused CSI
plugins. Plugins are self-cleaning when the last allocation for a
plugin is stopped, but this feature will cover any missing edge cases
and ensure that upgrades from 0.11.0 and 0.11.1 get any stray plugins
cleaned up.
This changeset adds a subsystem to run on the leader, similar to the
deployment watcher or node drainer. The `Watcher` performs a blocking
query on updates to the `CSIVolumes` table and triggers reaping of
volume claims.
This will avoid tying up scheduling workers by immediately sending
volume claim workloads into their own loop, rather than blocking the
scheduling workers in the core GC job doing things like talking to CSI
controllers
The volume watcher is enabled on leader step-up and disabled on leader
step-down.
The volume claim GC mechanism now makes an empty claim RPC for the
volume to trigger an index bump. That in turn unblocks the blocking
query in the volume watcher so it can assess which claims can be
released for a volume.
This is from a merge conflict resolution that went the wrong direction.
I assumed the block had been added, but really it had been removed. Now,
it is removed once again.
Nomad jobs may be configured with a TaskGroup which contains a Service
definition that is Consul Connect enabled. These service definitions end
up establishing a Consul Connect Proxy Task (e.g. envoy, by default). In
the case where Consul ACLs are enabled, a Service Identity token is required
for these tasks to run & connect, etc. This changeset enables the Nomad Server
to recieve RPC requests for the derivation of SI tokens on behalf of instances
of Consul Connect using Tasks. Those tokens are then relayed back to the
requesting Client, which then injects the tokens in the secrets directory of
the Task.
Enable any Server to lookup the unique ClusterID. If one has not been
generated, and this node is the leader, generate a UUID and attempt to
apply it through raft.
The value is not yet used anywhere in this changeset, but is a prerequisite
for gh-6701.
Fixes a deadlock in leadership handling if leadership flapped.
Raft propagates leadership transition to Nomad through a NotifyCh channel.
Raft blocks when writing to this channel, so channel must be buffered or
aggressively consumed[1]. Otherwise, Raft blocks indefinitely in `raft.runLeader`
until the channel is consumed[1] and does not move on to executing follower
related logic (in `raft.runFollower`).
While Raft `runLeader` defer function blocks, raft cannot process any other
raft operations. For example, `run{Leader|Follower}` methods consume
`raft.applyCh`, and while runLeader defer is blocked, all raft log applications
or config lookup will block indefinitely.
Sadly, `leaderLoop` and `establishLeader` makes few Raft calls!
`establishLeader` attempts to auto-create autopilot/scheduler config [3]; and
`leaderLoop` attempts to check raft configuration [4]. All of these calls occur
without a timeout.
Thus, if leadership flapped quickly while `leaderLoop/establishLeadership` is
invoked and hit any of these Raft calls, Raft handler _deadlock_ forever.
Depending on how many times it flapped and where exactly we get stuck, I suspect
it's possible to get in the following case:
* Agent metrics/stats http and RPC calls hang as they check raft.Configurations
* raft.State remains in Leader state, and server attempts to handle RPC calls
(e.g. node/alloc updates) and these hang as well
As we create goroutines per RPC call, the number of goroutines grow over time
and may trigger a out of memory errors in addition to missed updates.
[1] d90d6d6bda/config.go (L190-L193)
[2] d90d6d6bda/raft.go (L425-L436)
[3] 2a89e47746/nomad/leader.go (L198-L202)
[4] 2a89e47746/nomad/leader.go (L877)
Here, we ensure that when leader only responds to RPC calls when state
store is up to date. At leadership transition or launch with restored
state, the server local store might not be caught up with latest raft
logs and may return a stale read.
The solution here is to have an RPC consistency read gate, enabled when
`establishLeadership` completes before we respond to RPC calls.
`establishLeadership` is gated by a `raft.Barrier` which ensures that
all prior raft logs have been applied.
Conversely, the gate is disabled when leadership is lost.
This is very much inspired by https://github.com/hashicorp/consul/pull/3154/files
Enterprise only.
Disable preemption for service and batch jobs by default.
Maintain backward compatibility in a x.y.Z release. Consider switching
the default for new clusters in the future.
Tim Gross