This ensures that server-to-server streaming RPC calls use the tls
wrapped connections.
Prior to this, `streamingRpcImpl` function uses tls for setting header
and invoking the rpc method, but returns unwrapped tls connection.
Thus, streaming writes fail with tls errors.
This tls streaming bug existed since 0.8.0[1], but PR #5654[2]
exacerbated it in 0.9.2. Prior to PR #5654, nomad client used to
shuffle servers at every heartbeat -- `servers.Manager.setServers`[3]
always shuffled servers and was called by heartbeat code[4]. Shuffling
servers meant that a nomad client would heartbeat and establish a
connection against all nomad servers eventually. When handling
streaming RPC calls, nomad servers used these local connection to
communicate directly to the client. The server-to-server forwarding
logic was left mostly unexercised.
PR #5654 means that a nomad client may connect to a single server only
and caused the server-to-server forward streaming RPC code to get
exercised more and unearthed the problem.
[1] https://github.com/hashicorp/nomad/blob/v0.8.0/nomad/rpc.go#L501-L515
[2] https://github.com/hashicorp/nomad/pull/5654
[3] https://github.com/hashicorp/nomad/blob/v0.9.1/client/servers/manager.go#L198-L216
[4] https://github.com/hashicorp/nomad/blob/v0.9.1/client/client.go#L1603
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
Rename SnapshotAfter to SnapshotMinIndex. The old name was not
technically accurate. SnapshotAtOrAfter is more accurate, but wordy and
still lacks context about what precisely it is at or after (the index).
SnapshotMinIndex was chosen as it describes the action (snapshot), a
constraint (minimum), and the object of the constraint (index).
The previous commit prevented evaluating plans against a state snapshot
which is older than the snapshot at which the plan was created. This is
correct and prevents failures trying to retrieve referenced objects that
may not exist until the plan's snapshot. However, this is insufficient
to guarantee consistency if the following events occur:
1. P1, P2, and P3 are enqueued with snapshot @ 100
2. Leader evaluates and applies Plan P1 with snapshot @ 100
3. Leader evaluates Plan P2 with snapshot+P1 @ 100
4. P1 commits @ 101
4. Leader evaluates applies Plan P3 with snapshot+P2 @ 100
Since only the previous plan is optimistically applied to the state
store, the snapshot used to evaluate a plan may not contain the N-2
plan!
To ensure plans are evaluated and applied serially we must consider all
previous plan's committed indexes when evaluating further plans.
Therefore combined with the last PR, the minimum index at which to
evaluate a plan is:
min(previousPlanResultIndex, plan.SnapshotIndex)
Plan application should use a state snapshot at or after the Raft index
at which the plan was created otherwise it risks being rejected based on
stale data.
This commit adds a Plan.SnapshotIndex which is set by workers when
submitting plan. SnapshotIndex is set to the Raft index of the snapshot
the worker used to generate the plan.
Plan.SnapshotIndex plays a similar role to PlanResult.RefreshIndex.
While RefreshIndex informs workers their StateStore is behind the
leader's, SnapshotIndex is a way to prevent the leader from using a
StateStore behind the worker's.
Plan.SnapshotIndex should be considered the *lower bound* index for
consistently handling plan application.
Plans must also be committed serially, so Plan N+1 should use a state
snapshot containing Plan N. This is guaranteed for plans *after* the
first plan after a leader election.
The Raft barrier on leader election ensures the leader's statestore has
caught up to the log index at which it was elected. This guarantees its
StateStore is at an index > lastPlanIndex.
- updated region in job metadata that gets persisted to nomad datastore
- fixed many unrelated unit tests that used an invalid region value
(they previously passed because hcl wasn't getting picked up and
the job would default to global region)
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.
Revert plan_apply.go changes from #5411
Since non-Command Raft messages do not update the StateStore index,
SnapshotAfter may unnecessarily block and needlessly fail in idle
clusters where the last Raft message is a non-Command message.
This is trivially reproducible with the dev agent and a job that has 2
tasks, 1 of which fails.
The correct logic would be to SnapshotAfter the previous plan's index to
ensure consistency. New clusters or newly elected leaders will not have
a previous plan, so the index the leader was elected should be used
instead.
Fix a case where `node.StatusUpdatedAt` was manipulated directly in
memory.
This ensures that StatusUpdatedAt is set in raft layer, and ensures that
the field is updated when node drain/eligibility is updated too.
Previous commit could introduce a deadlock if the capacityChangeCh was
full and the receiving side exited before freeing a slot for the sending
side could send. Flush would then block forever waiting to acquire the
lock just to throw the pending update away.
The race is around getting/setting the chan field, not chan operations,
so only lock around getting the chan field.
I assume the mutex was being released before sending on capacityChangeCh
to avoid blocking in the critical section, but:
1. This is race.
2. capacityChangeCh has a *huge* buffer (8096). If it's full things
already seem Very Bad, and a little backpressure seems appropriate.
This fixes a bug in the state store during plan apply. When
denormalizing preempted allocations it incorrectly set the preemptor's
job during the update. This eventually causes a panic downstream in the
client. Added a test assertion that failed before and passes after this fix
* master: (912 commits)
Update redirects.txt
Added redirect for Spark guide link
client: log when server list changes
docs: mention regression in task config validation
fix update to changelog
update CHANGELOG with datacenter config validation https://github.com/hashicorp/nomad/pull/5665
typo: "atleast" -> "at least"
implement nomad exec for rkt
docs: fixed typo
use pty/tty terminology similar to github.com/kr/pty
vendor github.com/kr/pty
drivers: implement streaming exec for executor based drivers
executors: implement streaming exec
executor: scaffolding for executor grpc handling
client: expose allocated memory per task
client improve a comment in updateNetworks
stalebot: Add 'thinking' as an exempt label (#5684)
Added Sparrow link
update links to use new canonical location
Add redirects for restructing done in GH-5667
...
Fixes#1795
Running restored allocations and pulling what allocations to run from
the server happen concurrently. This means that if a client is rebooted,
and has its allocations rescheduled, it may restart the dead allocations
before it contacts the server and determines they should be dead.
This commit makes tasks that fail to reattach on restore wait until the
server is contacted before restarting.
Currently when an evalbroker is disabled, it still recieves delayed
enqueues via log application in the fsm. This causes an ever growing
heap of evaluations that will never be drained, and can cause memory
issues in larger clusters, or when left running for an extended period
of time without a leader election.
This commit prevents the enqueuing of evaluations while we are
disabled, and relies on the leader restoreEvals routine to handle
reconciling state during a leadership transition.
Existing dequeues during an Enabled->Disabled broker state transition are
handled by the enqueueLocked function dropping evals.
Primarily a cleanup commit, however, currently there is a potential race
condition (that I'm not sure we've ever actually hit) during a flapping
SetEnabled/Disabled state where we may never correctly restart the eval
broker, if it was being called from multiple routines.
Our testing so far indicates that ugorji/go/codec maintains backward
compatiblity with the version we are using now, for purposes of Nomad
serialization.
Using latest ugorji/go allows us to get back to using upstream library,
get get the optimizations benefits in RPC paths (including code
generation optimizations).
ugorji/go introduced two significant changes:
* time binary format in debb8e2d2e. Setting `h.BasicHandle.TimeNotBuiltin = true` restores old behavior
* ugorji/go started honoring `json` tag as well:
v1.1.4 is the latest but has a bug in handling RawString that's fixed in
d09a80c1e0
.
`Index` is already included as part of `WriteMeta` embedding.
This is a backward compatible change: Clients never read the field; and
Server refernces to `EmitNodeEventsResponse.Index` would be using the
value in `WriteMeta`, which is consistent with other response structs.
This command will be used to send a signal to either a single task within an
allocation, or all of the tasks if <task-name> is omitted. If the sent signal
terminates the allocation, it will be treated as if the allocation has crashed,
rather than as if it was operator-terminated.
Signal validation is currently handled by the driver itself and nomad
does not attempt to restrict or validate them.
