Registration and restoring allocs don't share state or depend on each
other in any way (syncing allocs with servers is done outside of
registration).
Since restoring is synchronous, start the registration goroutine first.
For nodes with lots of allocs to restore or close to their heartbeat
deadline, this could be the difference between becoming "lost" or not.
Refactoring of 104067bc2b2002a4e45ae7b667a476b89addc162
Switch the MarkLive method for a chan that is closed by the client.
Thanks to @notnoop for the idea!
The old approach called a method on most existing ARs and TRs on every
runAllocs call. The new approach does a once.Do call in runAllocs to
accomplish the same thing with less work. Able to remove the gate
abstraction that did much more than was needed.
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.
Related to #4280
This PR adds
`client.allocs.<job>.<group>.<alloc>.<task>.memory.allocated` as a gauge
in bytes to metrics to ease calculating how close a task is to OOMing.
```
'nomad.client.allocs.memory.allocated.example.cache.6d98cbaf-d6bc-2a84-c63f-bfff8905a9d8.redis.rusty': 268435456.000
'nomad.client.allocs.memory.cache.example.cache.6d98cbaf-d6bc-2a84-c63f-bfff8905a9d8.redis.rusty': 5677056.000
'nomad.client.allocs.memory.kernel_max_usage.example.cache.6d98cbaf-d6bc-2a84-c63f-bfff8905a9d8.redis.rusty': 0.000
'nomad.client.allocs.memory.kernel_usage.example.cache.6d98cbaf-d6bc-2a84-c63f-bfff8905a9d8.redis.rusty': 0.000
'nomad.client.allocs.memory.max_usage.example.cache.6d98cbaf-d6bc-2a84-c63f-bfff8905a9d8.redis.rusty': 8908800.000
'nomad.client.allocs.memory.rss.example.cache.6d98cbaf-d6bc-2a84-c63f-bfff8905a9d8.redis.rusty': 876544.000
'nomad.client.allocs.memory.swap.example.cache.6d98cbaf-d6bc-2a84-c63f-bfff8905a9d8.redis.rusty': 0.000
'nomad.client.allocs.memory.usage.example.cache.6d98cbaf-d6bc-2a84-c63f-bfff8905a9d8.redis.rusty': 8208384.000
```
* client: was not using up-to-date client state in determining which alloc count towards allocated resources
* Update client/client.go
Co-Authored-By: cgbaker <cgbaker@hashicorp.com>
This fixes a confusing UX where a previously successful deployment's
healthy/unhealthy count would get updated if any allocations failed after
the deployment was already marked as successful.
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.
Currently, when logmon fails to reattach, we will retry reattachment to
the same pid until the task restart specification is exhausted.
Because we cannot clear hook state during error conditions, it is not
possible for us to signal to a future restart that it _shouldn't_
attempt to reattach to the plugin.
Here we revert to explicitly detecting reattachment seperately from a
launch of a new logmon, so we can recover from scenarios where a logmon
plugin has failed.
This is a net improvement over the current hard failure situation, as it
means in the most common case (the pid has gone away), we can recover.
Other reattachment failure modes where the plugin may still be running
could potentially cause a duplicate process, or a subsequent failure to launch
a new plugin.
If there was a duplicate process, it could potentially cause duplicate
logging. This is better than a production workload outage.
If there was a subsequent failure to launch a new plugin, it would fail
in the same (retry until restarts are exhausted) as the current failure
mode.
This adds a `nomad alloc restart` command and api that allows a job operator
with the alloc-lifecycle acl to perform an in-place restart of a Nomad
allocation, or a given subtask.
Remove runLaunched tracking as Run is *always* called for killable
TaskRunners. TaskRunners which fail before Run can be called (during
NewTaskRunner or Restore) are not killable as they're never added to the
client's alloc map.
I chose to make them more of integration tests since there's a lot more
plumbing involved. The internal implementation details of how we craft
task envs can now change and these tests will still properly assert the
task runtime environment is setup properly.
Some of the context uses in TR hooks are useless (Killed during Stop
never seems meaningful).
None of the hooks are interruptable for graceful shutdown which is
unfortunate and probably needs fixing.
Builds upon earlier commit that cleans up restored handles of terminal
allocs by also emitting terminated events and calling exited hooks when
appropriate.
The test is sadly quite complicated and peeks into things (logmon's
reattach config) AR doesn't normally have access to.
However, I couldn't find another way of asserting logmon got cleaned up
without resorting to smaller unit tests. Smaller unit tests risk
re-implementing dependencies in an unrealistic way, so I opted for an
ugly integration test.
This commit is a significant change. TR.Run is now always executed, even
for terminal allocations. This was changed to allow TR.Run to cleanup
(run stop hooks) if a handle was recovered.
This is intended to handle the case of Nomad receiving a
DesiredStatus=Stop allocation update, persisting it, but crashing before
stopping AR/TR.
The commit also renames task runner hook data as it was very easy to
accidently set state on Requests instead of Responses using the old
field names.
This code chooses to be conservative as opposed to optimal: when failing
to reattach to logmon simply return a recoverable error instead of
immediately trying to restart logmon.
The recoverable error will cause the task's restart policy to be
applied and a new logmon will be launched upon restart.
Trying to do the optimal approach of simply starting a new logmon
requires error string comparison and should be tested against a task
actively logging to assert the behavior (are writes blocked? dropped?).
There were multiple bugs here:
1. Reattach unmarshalling always returned an error because you can't
unmarshal into a nil pointer.
2. The hook data wasn't being saved because it was put on the request
struct, not the response struct.
3. The plugin configuration should only have reattach *or* a command
set. Not both.
4. Setting Done=true meant the hook was never re-run on agent restart so
reattaching was never attempted.
Track the download status of each artifact independently so that if only
one of many artifacts fails to download, completed artifacts aren't
downloaded again.
0.9.0beta2 contains a regression where artifact download errors would
not cause a task restart and instead immediately fail the task.
This restores the pre-0.9 behavior of retrying all artifact errors and
adds missing tests.
Fixes an issue where if a task was restarted after restating the client,
the task dir environment variables would not be populated. This PR fixes
this for both upgrades from 0.8.X and for normal 0.9 restarts.
Added ability to adjust the number of events the TaskRunner keeps as
there's no way to observe all events otherwise.
Task events differ slightly from 0.8 because 0.9 emits Terminated every
time a task exits instead of only when it exits on its own (not due to
restart or kill).
0.9 does not emit Killing/Killed for restarts like 0.8 which seems fine
as `Restart Signaled/Terminated/Restarting` is more descriptive.
Original v0.8 events emitted:
```
expected := []string{
"Received",
"Task Setup",
"Started",
"Restart Signaled",
"Killing",
"Killed",
"Restarting",
"Started",
"Restart Signaled",
"Killing",
"Killed",
"Restarting",
"Started",
"Restart Signaled",
"Killing",
"Killed",
"Not Restarting",
}
```
v0.9.0-dev started emitting a Terminated event every time a task process
exited. While this wasn't true in previous versions, it's a useful task
event because it's the only place for job operators to view the task's
exit code.
This behavior is asserted in the e2e/taskevents tests.
Track current memory usage, `memory.usage_in_bytes`, in addition to
`memory.max_memory_usage_in_bytes` and friends. This number is closer
what Docker reports.
Related to https://github.com/hashicorp/nomad/issues/5165 .
plugins/driver: update driver interface to support streaming stats
client/tr: use streaming stats api
TODO:
* how to handle errors and closed channel during stats streaming
* prevent tight loop if Stats(ctx) returns an error
drivers: update drivers TaskStats RPC to handle streaming results
executor: better error handling in stats rpc
docker: better control and error handling of stats rpc
driver: allow stats to return a recoverable error
This PR improves how killing a task is handled. Before the kill function
directly orchestrated the killing and was only valid while the task was
running. The new behavior is to mark the desired state and wait for the
task runner to converge to that state.
We were just emitting Killed/Terminated events before. In v0.8 we
emitted Killing/Killed, but lacked Terminated when explicitly stopping
a task. This change makes it so Terminated is always included, whether
explicitly stopping a task or it exiting on its own.
New output:
2019-01-04T14:58:51-08:00 Killed Task successfully killed
2019-01-04T14:58:51-08:00 Terminated Exit Code: 130, Signal: 2
2019-01-04T14:58:51-08:00 Killing Sent interrupt
2019-01-04T14:58:51-08:00 Leader Task Dead Leader Task in Group dead
2019-01-04T14:58:49-08:00 Started Task started by client
2019-01-04T14:58:49-08:00 Task Setup Building Task Directory
2019-01-04T14:58:49-08:00 Received Task received by client
Old (v0.8.6) output:
2019-01-04T22:14:54Z Killed Task successfully killed
2019-01-04T22:14:54Z Killing Sent interrupt. Waiting 5s before force killing
2019-01-04T22:14:54Z Leader Task Dead Leader Task in Group dead
2019-01-04T22:14:53Z Started Task started by client
2019-01-04T22:14:53Z Task Setup Building Task Directory
2019-01-04T22:14:53Z Received Task received by client
Simplify allocDir.Build() function to avoid depending on client/structs,
and remove a parameter that's always set to `false`.
The motivation here is to avoid a dependency cycle between
drivers/cstructs and alloc_dir.
The driver manager is modeled after the device manager and is started by the client.
It's responsible for handling driver lifecycle and reattachment state, as well as
processing the incomming fingerprint and task events from each driver. The mananger
exposes a method for registering event handlers for task events that is used by the
task runner to update the server when a task has been updated with an event.
Since driver fingerprinting has been implemented by the driver manager, it is no
longer needed in the fingerprint mananger and has been removed.
The RestartCount is not really suitable for use as a source of
uniqueness within task invocations as it is not monotonic, and interacts
with the restart stanza in a users config, so conflates restarts due to
task failures, with restarts due to enviromental changes, such as consul
template or vault secrets changing.
Here we instead use a substring from a uuid, which is more random than
we strictly need, but is nicer than rolling our own random string
generator here.
This creates a new buffered channel and goroutine on the allocrunner for
serializing updates to allocations. This allows us to take updates off
the routine that is used from processing updates from the server,
without having complicated machinery for tracking update lifetimes, or
other external synchronization.
This results in a nice performance improvement and signficantly better
throughput on batch changes such as preempting a large number of jobs
for a larger placement.
This commit reduces the locking required to shutdown or destroy
allocrunners, and allows parallel shutdown and destroy of allocrunners during
shutdown.
When starting an allocation that is preempting other allocs, we create a
new group allocation watcher, and then wait for the allocations to
terminate in the allocation PreRun hooks.
If there's no preempted allocations, then we simply provide a
NoopAllocWatcher.
As part of deprecating legacy drivers, we're moving the env package to a
new drivers/shared tree, as it is used by the modern docker and rkt
driver packages, and is useful for 3rd party plugins.
Looking at NewTaskRunner I'm unsure whether TaskRunner.TaskResources
(from which req.TaskResources is set) is intended to be nil at times or
if the TODO in NewTaskRunner is intended to ensure it is always non-nil.
The old approach was incomplete. Hook env vars are now:
* persisted and restored between agent restarts
* deterministic (LWW if 2 hooks set the same key)
This PR introduces a device hook that retrieves the device mount
information for an allocation. It also updates the computed node class
computation to take into account devices.
TODO Fix the task runner unit test. The environment variable is being
lost even though it is being properly set in the prestart hook.
We were incorrectly returning a 0 duration to the taskrunner when
determining when a task should restart. This would cause tasks to be
restarted immediately, ignoring the restart {} stanza in a users
configuration.
This commit causes us to return the restart duration to the task runner
so it may correctly delay further execution.
This change makes few compromises:
* Looks up the devices associated with tasks at look up time. Given
that `nomad alloc status` is called rarely generally (compared to stats
telemetry and general job reporting), it seems fine. However, the
lookup overhead grows bounded by number of `tasks x total-host-devices`,
which can be significant.
* `client.Client` performs the task devices->statistics lookup. It
passes self to alloc/task runners so they can look up the device statistics
allocated to them.
* Currently alloc/task runners are responsible for constructing the
entire RPC response for stats
* The alternatives for making task runners device statistics aware
don't seem appealing (e.g. having task runners contain reference to hostStats)
* On the alloc aggregation resource usage, I did a naive merging of task device statistics.
* Personally, I question the value of such aggregation, compared to
costs of struct duplication and bloating the response - but opted to be
consistent in the API.
* With naive concatination, device instances from a single device group used by separate tasks in the alloc, would be aggregated in two separate device group statistics.
For lifecycle operations such as Restart and Kill, the client should not
expect driver plugins to be well behaved and close their waitCh on
context cancelation. Always wait on the passed in context as well as the
waitCh.
* Migrated all of the old leader task tests and got them passing
* Refactor and consolidate task killing code in AR to always kill leader
tasks first
* Fixed lots of issues with state restoring
* Fixed deadlock in AR.Destroy if AR.Run had never been called
* Added a new in memory statedb for testing
* Stopping an alloc is implemented via Updates but update hooks are
*not* run.
* Destroying an alloc is a best effort cleanup.
* AllocRunner destroy hooks implemented.
* Disk migration and blocking on a previous allocation exiting moved to
its own package to avoid cycles. Now only depends on alloc broadcaster
instead of also using a waitch.
* AllocBroadcaster now only drops stale allocations and always keeps the
latest version.
* Made AllocDir safe for concurrent use
Lots of internal contexts that are currently unused. Unsure if they
should be used or removed.
Saves a tiny bit of cpu and some IO. Sadly doesn't prevent all IO on
duplicate writes as the transactions are still created and committed.
$ go test -bench=. -benchmem
goos: linux
goarch: amd64
pkg: github.com/hashicorp/nomad/helper/boltdd
BenchmarkWriteDeduplication_On-4 500 4059591 ns/op 23736 B/op 56 allocs/op
BenchmarkWriteDeduplication_Off-4 300 4115319 ns/op 25942 B/op 55 allocs/op
Not setting the host name led the Go HTTP client to expect a certificate
with a DNS-resolvable name. Since Nomad uses `${role}.${region}.nomad`
names ephemeral dir migrations were broken when TLS was enabled.
Added an e2e test to ensure this doesn't break again as it's very
difficult to test and the TLS configuration is very easy to get wrong.
