* scheduler: stopped-yet-running allocs are still running
* scheduler: test new stopped-but-running logic
* test: assert nonoverlapping alloc behavior
Also add a simpler Wait test helper to improve line numbers and save few
lines of code.
* docs: tried my best to describe #10446
it's not concise... feedback welcome
* scheduler: fix test that allowed overlapping allocs
* devices: only free devices when ClientStatus is terminal
* test: output nicer failure message if err==nil
Co-authored-by: Mahmood Ali <mahmood@hashicorp.com>
Co-authored-by: Michael Schurter <mschurter@hashicorp.com>
Log lines which include an error should use the full term "error"
as the context key. This provides consistency across the codebase
and avoids a Go style which operators might not be aware of.
Fixes#13505
This fixes#13505 by treating reserved_ports like we treat a lot of jobspec settings: merging settings from more global stanzas (client.reserved.reserved_ports) "down" into more specific stanzas (client.host_networks[].reserved_ports).
As discussed in #13505 there are other options, and since it's totally broken right now we have some flexibility:
Treat overlapping reserved_ports on addresses as invalid and refuse to start agents. However, I'm not sure there's a cohesive model we want to publish right now since so much 0.9-0.12 compat code still exists! We would have to explain to folks that if their -network-interface and host_network addresses overlapped, they could only specify reserved_ports in one place or the other?! It gets ugly.
Use the global client.reserved.reserved_ports value as the default and treat host_network[].reserverd_ports as overrides. My first suggestion in the issue, but @groggemans made me realize the addresses on the agent's interface (as configured by -network-interface) may overlap with host_networks, so you'd need to remove the global reserved_ports from addresses shared with a shared network?! This seemed really confusing and subtle for users to me.
So I think "merging down" creates the most expressive yet understandable approach. I've played around with it a bit, and it doesn't seem too surprising. The only frustrating part is how difficult it is to observe the available addresses and ports on a node! However that's a job for another PR.
Stream snapshot to FSM when restoring from archive
The `RestoreFromArchive` helper decompresses the snapshot archive to a
temporary file before reading it into the FSM. For large snapshots
this performs a lot of disk IO. Stream decompress the snapshot as we
read it, without first writing to a temporary file.
Add bexpr filters to the `RestoreFromArchive` helper.
The operator can pass these as `-filter` arguments to `nomad operator
snapshot state` (and other commands in the future) to include only
desired data when reading the snapshot.
As a performance optimization in the scheduler, feasibility checks
that apply to an entire class are only checked once for all nodes of
that class. Other feasibility checks are "available" checks because
they rely on more ephemeral characteristics and don't contribute to
the hash for the node class. This currently includes only CSI.
We have a separate fast path for "available" checks when the node has
already been marked eligible on the basis of class. This fast path has
a bug where it returns early rather than continuing the loop. This
causes the entire task group to be rejected.
Fix the bug by not returning early in the fast path and instead jump
to the top of the loop like all the other code paths in this method.
Includes a new test exercising topology at whole-scheduler level and a
fix for an existing test that should've caught this previously.
In the reconciler's filtering for tainted nodes, we use whether the
server supports disconnected clients as a gate to a bunch of our
logic, but this doesn't account for cases where the job doesn't have
`max_client_disconnect`. The only real consequence of this appears to
be that allocs on disconnected nodes are marked "complete" instead of
"lost".
* planner: expose ServerMeetsMinimumVersion via Planner interface
* filterByTainted: add flag indicating disconnect support
* allocReconciler: accept and pass disconnect support flag
* tests: update dependent tests
CSI `CreateVolume` RPC is idempotent given that the topology,
capabilities, and parameters are unchanged. CSI volumes have many
user-defined fields that are immutable once set, and many fields that
are not user-settable.
Update the `Register` RPC so that updating a volume via the API merges
onto any existing volume without touching Nomad-controlled fields,
while validating it with the same strict requirements expected for
idempotent `CreateVolume` RPCs.
Also, clarify that this state store method is used for everything, not just
for the `Register` RPC.
The allocReconciler's computeGroup function contained a significant amount of inline logic that was difficult to understand the intent of. This commit extracts inline logic into the following intention revealing subroutines. It also includes updates to the function internals also aimed at improving maintainability and renames some existing functions for the same purpose. New or renamed functions include.
Renamed functions
- handleGroupCanaries -> cancelUnneededCanaries
- handleDelayedLost -> createLostLaterEvals
- handeDelayedReschedules -> createRescheduleLaterEvals
New functions
- filterAndStopAll
- initializeDeploymentState
- requiresCanaries
- computeCanaries
- computeUnderProvisionedBy
- computeReplacements
- computeDestructiveUpdates
- computeMigrations
- createDeployment
- isDeploymentComplete
The spread iterator can panic when processing an evaluation, resulting
in an unrecoverable state in the cluster. Whenever a panicked server
restarts and quorum is restored, the next server to dequeue the
evaluation will panic.
To trigger this state:
* The job must have `max_parallel = 0` and a `canary >= 1`.
* The job must not have a `spread` block.
* The job must have a previous version.
* The previous version must have a `spread` block and at least one
failed allocation.
In this scenario, the desired changes include `(place 1+) (stop
1+), (ignore n) (canary 1)`. Before the scheduler can place the canary
allocation, it tries to find out which allocations can be
stopped. This passes back through the stack so that we can determine
previous-node penalties, etc. We call `SetJob` on the stack with the
previous version of the job, which will include assessing the `spread`
block (even though the results are unused). The task group spread info
state from that pass through the spread iterator is not reset when we
call `SetJob` again. When the new job version iterates over the
`groupPropertySets`, it will get an empty `spreadAttributeMap`,
resulting in an unexpected nil pointer dereference.
This changeset resets the spread iterator internal state when setting
the job, logging with a bypass around the bug in case we hit similar
cases, and a test that panics the scheduler without the patch.
Processing an evaluation is nearly a pure function over the state
snapshot, but we randomly shuffle the nodes. This means that
developers can't take a given state snapshot and pass an evaluation
through it and be guaranteed the same plan results.
But the evaluation ID is already random, so if we use this as the seed
for shuffling the nodes we can greatly reduce the sources of
non-determinism. Unfortunately golang map iteration uses a global
source of randomness and not a goroutine-local one, but arguably
if the scheduler behavior is impacted by this, that's a bug in the
iteration.
If processing a specific evaluation causes the scheduler (and
therefore the entire server) to panic, that evaluation will never
get a chance to be nack'd and cleared from the state store. It will
get dequeued by another scheduler, causing that server to panic, and
so forth until all servers are in a panic loop. This prevents the
operator from intervening to remove the evaluation or update the
state.
Recover the goroutine from the top-level `Process` methods for each
scheduler so that this condition can be detected without panicking the
server process. This will lead to a loop of recovering the scheduler
goroutine until the eval can be removed or nack'd, but that's much
better than taking a downtime.
* csi: resolve invalid claim states on read
It's currently possible for CSI volumes to be claimed by allocations
that no longer exist. This changeset asserts a reasonable state at
the state store level by registering these nil allocations as "past
claims" on any read. This will cause any pass through the periodic GC
or volumewatcher to trigger the unpublishing workflow for those claims.
* csi: make feasibility check errors more understandable
When the feasibility checker finds we have no free write claims, it
checks to see if any of those claims are for the job we're currently
scheduling (so that earlier versions of a job can't block claims for
new versions) and reports a conflict if the volume can't be scheduled
so that the user can fix their claims. But when the checker hits a
claim that has a GCd allocation, the state is recoverable by the
server once claim reaping completes and no user intervention is
required; the blocked eval should complete. Differentiate the
scheduler error produced by these two conditions.
* csi: resolve invalid claim states on read
It's currently possible for CSI volumes to be claimed by allocations
that no longer exist. This changeset asserts a reasonable state at
the state store level by registering these nil allocations as "past
claims" on any read. This will cause any pass through the periodic GC
or volumewatcher to trigger the unpublishing workflow for those claims.
* csi: make feasibility check errors more understandable
When the feasibility checker finds we have no free write claims, it
checks to see if any of those claims are for the job we're currently
scheduling (so that earlier versions of a job can't block claims for
new versions) and reports a conflict if the volume can't be scheduled
so that the user can fix their claims. But when the checker hits a
claim that has a GCd allocation, the state is recoverable by the
server once claim reaping completes and no user intervention is
required; the blocked eval should complete. Differentiate the
scheduler error produced by these two conditions.
This PR exposes the following existing`consul-template` configuration options to Nomad jobspec authors in the `{job.group.task.template}` stanza.
- `wait`
It also exposes the following`consul-template` configuration to Nomad operators in the `{client.template}` stanza.
- `max_stale`
- `block_query_wait`
- `consul_retry`
- `vault_retry`
- `wait`
Finally, it adds the following new Nomad-specific configuration to the `{client.template}` stanza that allows Operators to set bounds on what `jobspec` authors configure.
- `wait_bounds`
Co-authored-by: Tim Gross <tgross@hashicorp.com>
Co-authored-by: Michael Schurter <mschurter@hashicorp.com>
* Fixed name of `nomad.scheduler.allocs.reschedule` metric
* Added new metrics to metrics reference documentation
* Expanded definitions of "waiting" metrics
* Changelog entry for #10236 and #10237
Adds a package `scheduler/benchmarks` with some examples of profiling
and benchmarking the scheduler, along with helpers for loading
real-world data for profiling.
This tooling comes out of work done for #11712. These test benchmarks
have not been added to CI because these particular profiles are mostly
examples and the runs will add an excessive amount of time to CI runs
for code that rarely changes in a way that has any chance of impacting
performance.
When the scheduler picks a node for each evaluation, the
`LimitIterator` provides at most 2 eligible nodes for the
`MaxScoreIterator` to choose from. This keeps scheduling fast while
producing acceptable results because the results are binpacked.
Jobs with a `spread` block (or node affinity) remove this limit in
order to produce correct spread scoring. This means that every
allocation within a job with a `spread` block is evaluated against
_all_ eligible nodes. Operators of large clusters have reported that
jobs with `spread` blocks that are eligible on a large number of nodes
can take longer than the nack timeout to evaluate (60s). Typical
evaluations are processed in milliseconds.
In practice, it's not necessary to evaluate every eligible node for
every allocation on large clusters, because the `RandomIterator` at
the base of the scheduler stack produces enough variation in each pass
that the likelihood of an uneven spread is negligible. Note that
feasibility is checked before the limit, so this only impacts the
number of _eligible_ nodes available for scoring, not the total number
of nodes.
This changeset sets the iterator limit for "large" `spread` block and
node affinity jobs to be equal to the number of desired
allocations. This brings an example problematic job evaluation down
from ~3min to ~10s. The included tests ensure that we have acceptable
spread results across a variety of large cluster topologies.
In the system scheduler, if a subset of clients are filtered by class,
we hit a code path where the `AllocMetric` has been copied, but the
`Copy` method does not instantiate the various maps. This leads to an
assignment to a nil map. This changeset ensures that the maps are
non-nil before continuing.
The `Copy` method relies on functions in the `helper` package that all
return nil slices or maps when passed zero-length inputs. This
changeset to fix the panic bug intentionally defers updating those
functions because it'll have potential impact on memory usage. See
https://github.com/hashicorp/nomad/issues/11564 for more details.
This change modifies the Nomad job register and deregister RPCs to
accept an updated option set which includes eval priority. This
param is optional and override the use of the job priority to set
the eval priority.
In order to ensure all evaluations as a result of the request use
the same eval priority, the priority is shared to the
allocReconciler and deploymentWatcher. This creates a new
distinction between eval priority and job priority.
The Nomad agent HTTP API has been modified to allow setting the
eval priority on job update and delete. To keep consistency with
the current v1 API, job update accepts this as a payload param;
job delete accepts this as a query param.
Any user supplied value is validated within the agent HTTP handler
removing the need to pass invalid requests to the server.
The register and deregister opts functions now all for setting
the eval priority on requests.
The change includes a small change to the DeregisterOpts function
which handles nil opts. This brings the function inline with the
RegisterOpts.
The system scheduler should leave allocs on draining nodes as-is, but
stop node stop allocs on nodes that are no longer part of the job
datacenters.
Previously, the scheduler did not make the distinction and left system
job allocs intact if they are already running.
I've added a failing test first, which you can see in https://app.circleci.com/jobs/github/hashicorp/nomad/179661 .
Fixes https://github.com/hashicorp/nomad/issues/11373
Fix a bug where the scheduler may panic when preemption is enabled. The conditions are a bit complicated:
A job with higher priority that schedule multiple allocations that preempt other multiple allocations on the same node, due to port/network/device assignments.
The cause of the bug is incidental mutation of internal cached data. `RankedNode` computes and cache proposed allocations in https://github.com/hashicorp/nomad/blob/v1.1.6/scheduler/rank.go#L42-L53 . But scheduler then mutates the list to remove pre-emptable allocs in https://github.com/hashicorp/nomad/blob/v1.1.6/scheduler/rank.go#L293-L294, and `RemoveAllocs` mutates and sets the tail of cached slice with `nil`s triggering a nil-pointer derefencing case.
I fixed the issue by avoiding the mutation in `RemoveAllocs` - the micro-optimization there doesn't seem necessary.
Fixes https://github.com/hashicorp/nomad/issues/11342
