TestClientEndpoint_CreateNodeEvals flakes a bit but its output is very
confusing, as `structs.Evaluations` overrides GoString.
Here, we emit the entire struct of the evaluation, and hopefully we'll
figure out the problem the next time it happens
The `volumewatcher` has a 250ms batch window so claim updates will not
typically be large enough to risk exceeding the maximum raft message
size. But large jobs might have enough volume claims that this could
be a danger. Set a maximum batch size of 100 messages per
batch (roughly 33K), as a very conservative safety/robustness guard.
Co-authored-by: Chris Baker <1675087+cgbaker@users.noreply.github.com>
* volumewatcher: remove redundant log fields
The constructor for `volumeWatcher` already sets a `logger.With` that
includes the volume ID and namespace fields. Remove them from the
various trace logs.
* volumewatcher: advance state for controller already released
One way of bypassing client RPCs in testing is to set a claim status
to controller-detached, but this results in an incorrect claim state
when we checkpoint.
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.
In this changeset:
* If a Nomad client node is running both a controller and a node
plugin (which is a common case), then if only the controller or the
node is removed, the plugin was not being updated with the correct
counts.
* The existing test for plugin cleanup didn't go back to the state
store, which normally is ok but is complicated in this case by
denormalization which changes the behavior. This commit makes the
test more comprehensive.
* Set "controller required" when plugin has `PUBLISH_READONLY`. All
known controllers that support `PUBLISH_READONLY` also support
`PUBLISH_UNPUBLISH_VOLUME` but we shouldn't assume this.
* Only create plugins when the allocs for those plugins are
healthy. If we allow a plugin to be created for the first time when
the alloc is not healthy, then we'll recreate deleted plugins when
the job's allocs all get marked terminal.
* Terminal plugin alloc updates should cleanup the plugin. The client
fingerprint can't tell if the plugin is unhealthy intentionally (for
the case of updates or job stop). Allocations that are
server-terminal should delete themselves from the plugin and trigger
a plugin self-GC, the same as an unused node.
The nil-check here is left-over from an earlier approach that didn't
get merged. It doesn't do anything for us now as we can't ever pass it
`nil` and if we leave it in the `getVolume` call it guards will panic
anyways.
- track lastHeartbeat, the client local time of the last successful
heartbeat round trip
- track allocations with `stop_after_client_disconnect` configured
- trigger allocation destroy (which handles cleanup)
- restore heartbeat/killable allocs tracking when allocs are recovered from disk
- on client restart, stop those allocs after a grace period if the
servers are still partioned
We should only remove the `ReadAllocs`/`WriteAllocs` values for a
volume after the claim has entered the "ready to free"
state. The volume will eventually be released as expected. But
querying the volume API will show the volume is released before the
controller unpublish has finished and this can cause a race with
starting new jobs.
Test updates are to cover cases where we're dropping claims but not
running through the whole reaping process.
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.
The `CSIVolumeClaim` fields were added after 0.11.1, so claims made
before that may be missing the value. Repair this when we read the
volume out of the state store.
The `NodeID` field was added after 0.11.0, so we need to ensure it's
been populated during upgrades from 0.11.0.
Adds a `CSIVolumeClaim` type to be tracked as current and past claims
on a volume. Allows for a client RPC failure during node or controller
detachment without having to keep the allocation around after the
first garbage collection eval.
This changeset lays groundwork for moving the actual detachment RPCs
into a volume watching loop outside the GC eval.
The BinPackIter accounted for node reservations twice when scoring nodes
which could bias scores toward nodes with reservations.
Pseudo-code for previous algorithm:
```
proposed = reservedResources + sum(allocsResources)
available = nodeResources - reservedResources
score = 1 - (proposed / available)
```
The node's reserved resources are added to the total resources used by
allocations, and then the node's reserved resources are later
substracted from the node's overall resources.
The new algorithm is:
```
proposed = sum(allocResources)
available = nodeResources - reservedResources
score = 1 - (proposed / available)
```
The node's reserved resources are no longer added to the total resources
used by allocations.
My guess as to how this bug happened is that the resource utilization
variable (`util`) is calculated and returned by the `AllocsFit` function
which needs to take reserved resources into account as a basic
feasibility check.
To avoid re-calculating alloc resource usage (because there may be a
large number of allocs), we reused `util` in the `ScoreFit` function.
`ScoreFit` properly accounts for reserved resources by subtracting them
from the node's overall resources. However since `util` _also_ took
reserved resources into account the score would be incorrect.
Prior to the fix the added test output:
```
Node: reserved Score: 1.0000
Node: reserved2 Score: 1.0000
Node: no-reserved Score: 0.9741
```
The scores being 1.0 for *both* nodes with reserved resources is a good
hint something is wrong as they should receive different scores. Upon
further inspection the double accounting of reserved resources caused
their scores to be >1.0 and clamped.
After the fix the added test outputs:
```
Node: no-reserved Score: 0.9741
Node: reserved Score: 0.9480
Node: reserved2 Score: 0.8717
```
The field names within the structs representing the Connect proxy
definition were not the same (nomad/structs/ vs api/), causing the
values to be lost in translation for the 'nomad job inspect' command.
Since the field names already shipped in v0.11.0 we cannot simply
fix the names. Instead, use the json struct tag on the structs/ structs
to remap the name to match the publicly expose api/ package on json
encoding.
This means existing jobs from v0.11.0 will continue to work, and
the JSON API for job submission will remain backwards compatible.
Follow-up for a method missed in the refactor for #7688. The
`volAndPluginLookup` method is only ever called from the server's `CSI`
RPC and never the `ClientCSI` RPC, so move it into that scope.
The current design of `ClientCSI` RPC requires that callers in the
server know about the free-standing `nodeForControllerPlugin`
function. This makes it difficult to send `ClientCSI` RPC messages
from subpackages of `nomad` and adds a bunch of boilerplate to every
server-side caller of a controller RPC.
This changeset makes it so that the `ClientCSI` RPCs will populate and
validate the controller's client node ID if it hasn't been passed by
the caller, centralizing the logic of picking and validating
controller targets into the `nomad.ClientCSI` struct.
Before, if the sidecar_service stanza of a connect enabled service
was missing, the job submission would cause a panic in the nomad
agent. Since the panic was happening in the API handler the agent
itself continued running, but this change will the condition more
gracefully.
By fixing the `Copy` method, the API handler now returns the proper
error.
$ nomad job run foo.nomad
Error submitting job: Unexpected response code: 500 (1 error occurred:
* Task group api validation failed: 2 errors occurred:
* Missing tasks for task group
* Task group service validation failed: 1 error occurred:
* Service[0] count-api validation failed: 1 error occurred:
* Consul Connect must be native or use a sidecar service
The `Job.Deregister` call will block on the client CSI controller RPCs
while the alloc still exists on the Nomad client node. So we need to
make the volume claim reaping async from the `Job.Deregister`. This
allows `nomad job stop` to return immediately. In order to make this
work, this changeset changes the volume GC so that the GC jobs are on a
by-volume basis rather than a by-job basis; we won't have to query
the (possibly deleted) job at the time of volume GC. We smuggle the
volume ID and whether it's a purge into the GC eval ID the same way we
smuggled the job ID previously.
The CSI plugins uses the external volume ID for all operations, but
the Client CSI RPCs uses the Nomad volume ID (human-friendly) for the
mount paths. Pass the External ID as an arg in the RPC call so that
the unpublish workflows have it without calling back to the server to
find the external ID.
The controller CSI plugins need the CSI node ID (or in other words,
the storage provider's view of node ID like the EC2 instance ID), not
the Nomad node ID, to determine how to detach the external volume.
* nomad/state/state_store: error message copy/paste error
* nomad/structs/structs: add a VolumeEval to the JobDeregisterResponse
* nomad/job_endpoint: synchronously, volumeClaimReap on job Deregister
* nomad/core_sched: make volumeClaimReap available without a CoreSched
* nomad/job_endpoint: Deregister return early if the job is missing
* nomad/job_endpoint_test: job Deregistion is idempotent
* nomad/core_sched: conditionally ignore alloc status in volumeClaimReap
* nomad/job_endpoint: volumeClaimReap all allocations, even running
* nomad/core_sched_test: extra argument to collectClaimsToGCImpl
* nomad/job_endpoint: job deregistration is not idempotent
I hypothesize that the flakiness in rolling update is due to shutting
down s3 server before s4 is properly added as a voter.
The chain of the flakiness is as follows:
1. Bootstrap with s1, s2, s3
2. Add s4
3. Wait for servers to register with 3 voting peers
* But we already have 3 voters (s1, s2, and s3)
* s4 is added as a non-voter in Raft v3 and must wait until autopilot promots it
4. Test proceeds without s4 being a voter
5. s3 shutdown
6. cluster changes stall due to leader election and too many pending configuration
changes (e.g. removing s3 from raft, promoting s4).
Here, I have the test wait until s4 is marked as a voter before shutting
down s3, so we don't have too many configuration changes at once.
In https://circleci.com/gh/hashicorp/nomad/57092, I noticed the
following events:
```
TestAutopilot_RollingUpdate: autopilot_test.go:204: adding server s4
TestAutopilot_RollingUpdate: testlog.go:34: 2020-04-03T20:08:19.789Z [INFO] nomad/serf.go:60: nomad: adding server: server="nomad-137.global (Addr: 127.0.0.1:9177) (DC: dc1)"
TestAutopilot_RollingUpdate: testlog.go:34: 2020-04-03T20:08:19.789Z [INFO] raft/raft.go:1018: nomad.raft: updating configuration: command=AddNonvoter server-id=c54b5bf4-1159-34f6-032d-56aefeb08425 server-addr=127.0.0.1:9177 servers="[{Suffrage:Voter ID:df01ba65-d1b2-17a9-f792-a4459b3a7c09 Address:127.0.0.1:9171} {Suffrage:Voter ID:c3337778-811e-2675-87f5-006309888387 Address:127.0.0.1:9173} {Suffrage:Voter ID:186d5e15-c473-e2b3-b5a4-3259a84e10ef Address:127.0.0.1:9169} {Suffrage:Nonvoter ID:c54b5bf4-1159-34f6-032d-56aefeb08425 Address:127.0.0.1:9177}]"
TestAutopilot_RollingUpdate: autopilot_test.go:218: shutting down server s3
TestAutopilot_RollingUpdate: testlog.go:34: 2020-04-03T20:08:19.797Z [INFO] raft/replication.go:456: nomad.raft: aborting pipeline replication: peer="{Nonvoter c54b5bf4-1159-34f6-032d-56aefeb08425 127.0.0.1:9177}"
TestAutopilot_RollingUpdate: autopilot_test.go:235: waiting for s4 to stabalize and be promoted
TestAutopilot_RollingUpdate: testlog.go:34: 2020-04-03T20:08:19.975Z [ERROR] raft/raft.go:1656: nomad.raft: failed to make requestVote RPC: target="{Voter c3337778-811e-2675-87f5-006309888387 127.0.0.1:9173}" error="dial tcp 127.0.0.1:9173: connect: connection refused"
TestAutopilot_RollingUpdate: retry.go:121: autopilot_test.go:241: don't want "c3337778-811e-2675-87f5-006309888387"
autopilot_test.go:241: didn't find map[c54b5bf4-1159-34f6-032d-56aefeb08425:true] in []raft.ServerID{"df01ba65-d1b2-17a9-f792-a4459b3a7c09", "186d5e15-c473-e2b3-b5a4-3259a84e10ef"}
```
Note how s3, c3337778, is present in the peers list in the final
failure, but s4, c54b5bf4, is added as a Nonvoter and isn't present in
the final peers list.
If a volume-claiming alloc stops and the CSI Node plugin that serves
that alloc's volumes is missing, there's no way for the allocrunner
hook to send the `NodeUnpublish` and `NodeUnstage` RPCs.
This changeset addresses this issue with a redesign of the client-side
for CSI. Rather than unmounting in the alloc runner hook, the alloc
runner hook will simply exit. When the server gets the
`Node.UpdateAlloc` for the terminal allocation that had a volume claim,
it creates a volume claim GC job. This job will made client RPCs to a
new node plugin RPC endpoint, and only once that succeeds, move on to
making the client RPCs to the controller plugin. If the node plugin is
unavailable, the GC job will fail and be requeued.
* nomad/state/state_store: enforce that the volume namespace exists
* nomad/csi_endpoint_test: a couple of broken namespaces now
* nomad/csi_endpoint_test: one more test
* nomad/node_endpoint_test: use structs.DefaultNamespace
* nomad/state/state_store_test: use DefaultNamespace
