**The Bug:**
You may have seen log lines like this when running 0.9.0-dev:
```
... client.alloc_runner.task_runner: some environment variables not available for rendering: ... keys="attr.driver.docker.volumes.enabled, attr.driver.docker.version, attr.driver.docker.bridge_ip, attr.driver.qemu.version"
```
Not only should we not be erroring on builtin driver attributes, but the
results were nondeterministic due to map iteration order!
The root cause is that we have an old root attribute for all drivers
like:
```
attr.driver.docker = "1"
```
When attributes were opaque variable names it was fine to also have
"nested" attributes like:
```
attr.driver.docker.version = "1.2.3"
```
However in the HCLv2 world the variable names are no longer opaque: they
form an object tree. The `docker` object can no longer both hold a value
(`"1"`) *and* nested attributes (`version = "1.2.3"`).
**The Fix:**
Since the old `attr.driver.<name> = "1"` attribues are useless for task
config interpolation, create a new precedence rule for creating the task
config evaluation context:
*Maps take precedence over primitives.*
This means `attr.driver.docker.version` will always take precedence over
`attr.driver.docker`. The results are determinstic and give users access
to the more useful metadata.
I made this a general precedence rule instead of special-casing driver
attrs because it seemed like better default behavior than spamming
WARNings to logs that were likely unactionable by users.
Unfortunately I don't know how to test these errors. As far as I can
tell they should only happen if there was a programming error in the
upgrade code or the underlying boltdb was corrupted somehow.
* Prefix task bucket with task- to prevent name conflicts
* Shorten device manager bucket name
* Remove commented out outdated var
* Update layout comment
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.
The assertion here is causing many spurious failures that aren't
actually relevant to the test itself.
We are tracking the cause for this failure independently, and it would
make more sense to have a dedicated test for clean shutdown.
Currently, there is a race condition between creating a taskrunner, and
updating node attributes via fingerprinting.
This is because the taskenv builder will try to iterate over the
clientconfig.Node.Attributes map, which can be concurrently updated by
the fingerprinting process, thus causing a panic.
This fixes that by providing a copy of the clientconfg to the
allocrunner inside the Read lock during config creation.
The allocLock is used to synchronize access to the alloc runner map, not
to ensure internal consistency of the alloc runners themselves. This
updates the updateAlloc process to avoid hanging on to an exclusive lock
of the map while applying changes to allocrunners themselves, as they
should be internally consistent.
This fixes a bug where any client allocation api will block during the
shutdown or updating of an allocrunner and its child taskrunners.
Fixes a bug where a driver health and attributes are never updated from
their initial status. If a driver started unhealthy, it may never go
into a healthy status.
Noticed few places where tests seem to block indefinitely and panic
after the test run reaches the test package timeout.
I intend to follow up with the proper fix later, but timing out is much
better than indefinitely blocking.
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.
The Group Alloc watcher is an implementation of a PrevAllocWatcher that
can wait for multiple previous allocs before terminating.
This is to be used when running an allocation that is preempting upstream
allocations, and thus only supports being ran with a local alloc watcher.
It also currently requires all of its child watchers to correctly handle
context cancellation. Should this be a problem, it should be fairly easy
to implement a replacement using channels rather than a waitgroup.
It obeys the PrevAllocWatcher interface for convenience, but it may be
better to extract Migration capabilities into a seperate interface for
greater clarity.
As of now, FileRotator uses bufio.Write under the hood to write data to
configured output file. Due to the way how bufio handles any occurred io
error - saves it into `err` variable never resetting it automatically -
any operation like `Write`, `Flush` etc will become a no-op, returning the very same,
saved error (eg. Out of disk space) even when the problem is fixed (eg. disk
space is available again).
That automatically means that FileRotator will stop writing any logs,
reporting the same error over and over again, even if it's no longer
valid.
This PR fixes it by resetting the bufio Writer, which resets any errors
and tries to write requested data.
IOPS have been modelled as a resource since Nomad 0.1 but has never
actually been detected and there is no plan in the short term to add
detection. This is because IOPS is a bit simplistic of a unit to define
the performance requirements from the underlying storage system. In its
current state it adds unnecessary confusion and can be removed without
impacting any users. This PR leaves IOPS defined at the jobspec parsing
level and in the api/ resources since these are the two public uses of
the field. These should be considered deprecated and only exist to allow
users to stop using them during the Nomad 0.9.x release. In the future,
there should be no expectation that the field will exist.
The previous integration test was broken during the client refactor, and
it seems to be some sort of race with state updating.
I'm going to try and construct a replacement test as part of work on
performance, but for now, the underlying behaviour is still being
tested.
This removes a cyclical dependency when importing client/structs from
dependencies of the plugin_loader, specifically, drivers. Due to
client/config also depending on the plugin_loader.
It also better reflects the ownership of fingerprint structs, as they
are fairly internal to the fingerprint manager.
The logging package is used by logmon and the legacy mock_driver. Because the
legacy drivers are going away, I'm moving it here to signify its actual
ownership.
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.
The group utility struct does not support asynchronously launched
goroutines (goroutines-inside-of-goroutines), so switch those uses to a
normal go call.
This means watchNodeUpdates and watchNodeEvents may not be shutdown when
Shutdown() exits. During nomad agent shutdown this does not matter.
During tests this means a test may leak those goroutines or be unable to
know when those goroutines have exited.
Since there's no runtime impact and these goroutines do not affect alloc
state syncing it seems ok to risk leaking them.
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.
In state values, we need to be able to distinguish between zero values
(e.g. `false`) and unset values (e.g. `nil`).
We can alternatively use protobuf `oneOf` and nested map to ensure
consistency of fields that are set together, but the golang
representation does not represent that well and introducing a mismatch
between representations. Thus, I opted not to use it.
Tests expect that as soon as eventer shuts down immediately on context
cancellations; but golang does not guarantee priority when multiple
pending channels are ready in a select statement.
The default job here contains some exec task config (for setting
command and args) that aren't used for mock driver. Now, the alloc
runner seems stricter about validating fields and errors on unexpected
fields.
Updating configs in tests so we can have an explicit task config
whenever driver is set explicitly.
Introduce a device manager that manages the lifecycle of device plugins
on the client. It fingerprints, collects stats, and forwards Reserve
requests to the correct plugin. The manager, also handles device plugins
failing and validates their output.
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
If ar.state.TaskStates has not been set, set it on the copy of ar.state.
That keeps ar.state manipulations in one location and allows AllocState
to only acquire read-locks.
Although the really exciting change is making WaitForRunning return the
allocations that it started. This should cut down test boilerplate
significantly.
The interesting decision in this commit was to expose AR's state and not
a fully materialized Allocation struct. AR.clientAlloc builds an Alloc
that contains the task state, so I considered simply memoizing and
exposing that method.
However, that would lead to AR having two awkwardly similar methods:
- Alloc() - which returns the server-sent alloc
- ClientAlloc() - which returns the fully materialized client alloc
Since ClientAlloc() could be memoized it would be just as cheap to call
as Alloc(), so why not replace Alloc() entirely?
Replacing Alloc() entirely would require Update() to immediately
materialize the task states on server-sent Allocs as there may have been
local task state changes since the server received an Alloc update.
This quickly becomes difficult to reason about: should Update hooks use
the TaskStates? Are state changes caused by TR Update hooks immediately
reflected in the Alloc? Should AR persist its copy of the Alloc? If so,
are its TaskStates canonical or the TaskStates on TR?
So! Forget that. Let's separate the static Allocation from the dynamic
AR & TR state!
- AR.Alloc() is for static Allocation access (often for the Job)
- AR.AllocState() is for the dynamic AR & TR runtime state (deployment
status, task states, etc).
If code needs to know the status of a task: AllocState()
If code needs to know the names of tasks: Alloc()
It should be very easy for a developer to reason about which method they
should call and what they can do with the return values.
Multiple receivers raced for the WaitResult when killing tasks which
could lead to a deadlock if the "wrong" receiver won.
Wrap handlers in an ugly little proxy to avoid this. At first I wanted
to push this into drivers, but the result is tied to the TR's handle
lifecycle -- not the lifecycle of an alloc or task.
"Ask forgiveness, not permission."
Instead of peaking at TaskStates (which are no longer updated on the
AR.Alloc() view of the world) to only read logs for running tasks, just
try to read the logs and improve the error handling if they don't exist.
This should make log streaming less dependent on AR/TR behavior.
Also fixed a race where the log streamer could exit before reading an
error. This caused no logs or errors to be displayed sometimes when an
error occurred.
Driver plugin framework to facilitate development of driver plugins.
Implementing plugins only need to implement the DriverPlugin interface.
The framework proxies this interface to the go-plugin GRPC interface generated
from the driver.proto spec.
A testing harness is provided to allow implementing drivers to test the full
lifecycle of the driver plugin. An example use:
func TestMyDriver(t *testing.T) {
harness := NewDriverHarness(t, &MyDiverPlugin{})
// The harness implements the DriverPlugin interface and can be used as such
taskHandle, err := harness.StartTask(...)
}
* client/executor: refactor client to remove interpolation
* executor: POC libcontainer based executor
* vendor: use hashicorp libcontainer fork
* vendor: add libcontainer/nsenter dep
* executor: updated executor interface to simplify operations
* executor: implement logging pipe
* logmon: new logmon plugin to manage task logs
* driver/executor: use logmon for log management
* executor: fix tests and windows build
* executor: fix logging key names
* executor: fix test failures
* executor: add config field to toggle between using libcontainer and standard executors
* logmon: use discover utility to discover nomad executable
* executor: only call libcontainer-shim on main in linux
* logmon: use seperate path configs for stdout/stderr fifos
* executor: windows fixes
* executor: created reusable pid stats collection utility that can be used in an executor
* executor: update fifo.Open calls
* executor: fix build
* remove executor from docker driver
* executor: Shutdown func to kill and cleanup executor and its children
* executor: move linux specific universal executor funcs to seperate file
* move logmon initialization to a task runner hook
* client: doc fixes and renaming from code review
* taskrunner: use shared config struct for logmon fifo fields
* taskrunner: logmon only needs to be started once per task
Updated to hclog.
It exposed fields that required an unexported lock to access. Created a
getter methodn instead. Only old allocrunner currently used this
feature.
* vendor: bump libcontainer and docker to remove Sirupsen imports
* vendor: fix bad vendoring of archive package
* vendor: fix api changes to cgroups in executor
* vendor: fix docker api changes
* vendor: update github.com/Azure/go-ansiterm to use non capitalized logrus import
* UpdateState: set state, append event, persist, update servers
* EmitEvent: append event, persist, update servers
* AppendEvent: append event, persist
AppendEvent may not even have to persist, but for the sake of
correctness I'm going with that for now.
* 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
