In #24658 we fixed a bug around client restarts where we would not assert
network namespaces existed and were properly configured when restoring
allocations. We introduced a call to the CNI `Check` method so that the plugins
could report correct config. But when we get an error from this call, we don't
log it unless the error is fatal. This makes it challenging to debug the case
where the initial check fails but we tear down the network and try again (as
described in #25510). Add a noisy log line here.
Ref: https://github.com/hashicorp/nomad/pull/24658
Ref: https://github.com/hashicorp/nomad/issues/25510
Fix the checking of the staging path against the mountRoot on the host
rather then checking against the containerMountPoint which (probably)
never exists on the host causing it to default back the the legacy
behaviour.
Some of our allocrunner hooks require a task environment for interpolating values based on the node or allocation. But several of the hooks accept an already-built environment or builder and then keep that in memory. Both of these retain a copy of all the node attributes and allocation metadata, which balloons memory usage until the allocation is GC'd.
While we'd like to look into ways to avoid keeping the allocrunner around entirely (see #25372), for now we can significantly reduce memory usage by creating the task environment on-demand when calling allocrunner methods, rather than persisting it in the allocrunner hooks.
In doing so, we uncover two other bugs:
* The WID manager, the group service hook, and the checks hook have to interpolate services for specific tasks. They mutated a taskenv builder to do so, but each time they mutate the builder, they write to the same environment map. When a group has multiple tasks, it's possible for one task to set an environment variable that would then be interpolated in the service definition for another task if that task did not have that environment variable. Only the service definition interpolation is impacted. This does not leak env vars across running tasks, as each taskrunner has its own builder.
To fix this, we move the `UpdateTask` method off the builder and onto the taskenv as the `WithTask` method. This makes a shallow copy of the taskenv with a deep clone of the environment map used for interpolation, and then overwrites the environment from the task.
* The checks hook interpolates Nomad native service checks only on `Prerun` and not on `Update`. This could cause unexpected deregistration and registration of checks during in-place updates. To fix this, we make sure we interpolate in the `Update` method.
I also bumped into an incorrectly implemented interface in the CSI hook. I've pulled that and some better guardrails out to https://github.com/hashicorp/nomad/pull/25472.
Fixes: https://github.com/hashicorp/nomad/issues/25269
Fixes: https://hashicorp.atlassian.net/browse/NET-12310
Ref: https://github.com/hashicorp/nomad/issues/25372
While working on #25373, I noticed that the CSI hook's `Destroy` method doesn't
match the interface, which means it never gets called. Because this method only
cancels any in-flight CSI requests, the only impact of this bug is that any CSI
RPCs that are in-flight when an alloc is GC'd on the client or a dev agent is
shut down won't be interrupted gracefully.
Fix the interface, but also make static assertions for all the allocrunner hooks
in the production code, so that you can make changes to interfaces and have
compile-time assistance in avoiding mistakes.
Ref: https://github.com/hashicorp/nomad/pull/25373
* test: use statedb factory
Swapping fields on Client after it has been created is a race.
* test: lock before checking heartbeat state
Fixes races
* test: fix races by copying fsm objects
A common source of data races in tests is when they insert a fixture
directly into memdb and then later mutate the object. Since objects in
the state store are readonly, any later mutation is a data race.
* test: lock when peeking at eval stats
* test: lock when peeking at serf state
* test: lock when looking at stats
* test: fix default eval broker state test
The test was not applying the config callback. In addition the test
raced against the configuration being applied. Waiting for the keyring
to be initialized resolved the race in my testing, but given the high
concurrency of the various leadership subsystems it's possible it may
still flake.
This change removes any blocking calls to destroyAllocRunner, which
caused nomad clients to block when running allocations in certain
scenarios. In addition, this change consolidates client GC by removing
the MakeRoomFor method, which is redundant to keepUsageBelowThreshold.
---------
Co-authored-by: Tim Gross <tgross@hashicorp.com>
Errors from `volume create` or `volume delete` only get logged by the client
agent, which may make it harder for volume authors to debug these tasks if they
are not also the cluster administrator with access to host logs.
Allow plugins to include an optional error message in their response. Because we
can't count on receiving this response (the error could come before the plugin
executes), we parse this message optimistically and include it only if
available.
Ref: https://hashicorp.atlassian.net/browse/NET-12087
When a CSI plugin is launched, we probe it until the csi_plugin.health_timeout
expires (by default 30s). But if the plugin never becomes healthy, we're not
restarting the task as documented.
Update the plugin supervisor to trigger a restart instead. We still exit the
supervisor loop at that point to avoid having the supervisor send probes to a
task that isn't running yet. This requires reworking the poststart hook to allow
the supervisor loop to be restarted when the task restarts.
In doing so, I identified that we weren't respecting the task kill context from
the post start hook, which would leave the supervisor running in the window
between when a task is killed because it failed and its stop hooks were
triggered. Combine the two contexts to make sure we stop the supervisor
whichever context gets closed first.
Fixes: https://github.com/hashicorp/nomad/issues/25293
Ref: https://hashicorp.atlassian.net/browse/NET-12264
The group level fields stop_after_client_disconnect,
max_client_disconnect, and prevent_reschedule_on_lost were deprecated in
Nomad 1.8 and replaced by field in the disconnect block. This change
removes any logic related to those deprecated fields.
---------
Co-authored-by: Tim Gross <tgross@hashicorp.com>
Before the fixes in #20165, the wait feature was disabled by
default. After these changes, it's always enabled, which - at
least on some platforms - leads to a significant increase in
load (5-7x).
This patch allows disabling the wait feature in the client
stanza of the configuration file by setting min and max to 0:
wait {
min = "0"
max = "0"
}
Per-template wait blocks in the task description still work like
one would expect.
When a node is fingerprinted, we calculate a "computed class" from a hash over a
subset of its fields and attributes. In the scheduler, when a given node fails
feasibility checking (before fit checking) we know that no other node of that
same class will be feasible, and we add the hash to a map so we can reject them
early. This hash cannot include any values that are unique to a given node,
otherwise no other node will have the same hash and we'll never save ourselves
the work of feasibility checking those nodes.
In #4390 we introduce the `nomad.advertise.address` attribute and in #19969 we
introduced `consul.dns.addr` attribute. Both of these are unique per node and
break the hash.
Additionally, we were not correctly filtering attributes out when checking if a
node escaped the class by not filtering for attributes that start with
`unique.`. The test for this introduced in #708 had an inverted assertion, which
allowed this to pass unnoticed since the early days of Nomad.
Ref: https://github.com/hashicorp/nomad/pull/708
Ref: https://github.com/hashicorp/nomad/pull/4390
Ref: https://github.com/hashicorp/nomad/pull/19969
The legacy workflow for Vault whereby servers were configured
using a token to provide authentication to the Vault API has now
been removed. This change also removes the workflow where servers
were responsible for deriving Vault tokens for Nomad clients.
The deprecated Vault config options used byi the Nomad agent have
all been removed except for "token" which is still in use by the
Vault Transit keyring implementation.
Job specification authors can no longer use the "vault.policies"
parameter and should instead use "vault.role" when not using the
default workload identity.
---------
Co-authored-by: Tim Gross <tgross@hashicorp.com>
Co-authored-by: Aimee Ukasick <aimee.ukasick@hashicorp.com>
In #24526 we updated Consul and Vault fingerprinting so that we no longer
periodically fingerprint. In #25102 we made it so that we fingerprint
periodically on start until the first fingerprint, in order to tolerate Consul
or Vault not being available on start. For clusters not running Consul, this
leads to a warn-level log every 15s. This same log exists for Vault, but Vault
support is opt-in via `vault.enable = true` whereas you have to manually disable
the fingerprinter for Consul.
Make it so that we only log a failed Consul fingerprint once per Consul
cluster. Reset the gate on this once we have a successful fingerprint, so that
we get the logs after a reload if Consul is unavailable.
Ref: https://github.com/hashicorp/nomad/pull/24526
Ref: https://github.com/hashicorp/nomad/pull/25102
Fixes: https://github.com/hashicorp/nomad/issues/25181
In #20165 we fixed a bug where a partially configured `client.template` retry
block would set any unset fields to nil instead of their default values. But
this patch introduced a regression in the default values, so we were now
defaulting to unlimited retries if the retry block was unset. Restore the
correct behavior and add better test coverage at both the config parsing and
template configuration code.
Ref: https://github.com/hashicorp/nomad/pull/20165
Ref: https://github.com/hashicorp/nomad/issues/23305#issuecomment-2643731565
In #24526 we updated the Consul and Vault fingerprints so that they are no
longer periodic. This fixed a problem that cluster admins reported where rolling
updates of Vault or Consul would cause a thundering herd of fingerprint updates
across the whole cluster.
But if Consul/Vault is not available during the initial fingerprint, it will
never get fingerprinted again. This is challenging for cluster updates and black
starts because the implicit service startup ordering may require
reloads. Instead, have the fingerprinter run periodically but mark that it has
made its first successful fingerprint of all Consul/Vault clusters. At that
point, we can skip further periodic updates. The `Reload` method will reset the
mark and allow the subsequent fingerprint to run normally.
Fixes: https://github.com/hashicorp/nomad/issues/25097
Ref: https://github.com/hashicorp/nomad/pull/24526
Ref: https://github.com/hashicorp/nomad/issues/24049
In #24650 we switched to using ephemeral state for CNI plugins, so that when a
host reboots and we lose all the allocations we don't end up trying to use IPs
we created in network namespaces we just destroyed. Unfortunately upgrade
testing missed that in a non-reboot scenario, the existing CNI state was being
used by plugins like the ipam plugin to hand out the "next available" IP
address. So with no state carried over, we might allocate new addresses that
conflict with existing allocations. (This can be avoided by draining the node
first.)
As a compatibility shim, copy the old CNI state directory to the new CNI state
directory during agent startup, if the new CNI state directory doesn't already
exist.
Ref: https://github.com/hashicorp/nomad/pull/24650
When a blocking query on the client hits a retryable error, we change the max
query time so that it falls within the `RPCHoldTimeout` timeout. But when the
retry succeeds we don't reset it to the original value.
Because the calls to `Node.GetClientAllocs` reuse the same request struct
instead of reallocating it, any retry will cause the agent to poll at a faster
frequency until the agent restarts. No other current RPC on the client has this
behavior, but we'll fix this in the `rpc` method rather than in the caller so
that any future users of the `rpc` method don't have to remember this detail.
Fixes: https://github.com/hashicorp/nomad/issues/25033
* Upgrade to using hashicorp/go-metrics@v0.5.4
This also requires bumping the dependencies for:
* memberlist
* serf
* raft
* raft-boltdb
* (and indirectly hashicorp/mdns due to the memberlist or serf update)
Unlike some other HashiCorp products, Nomads root module is currently expected to be consumed by others. This means that it needs to be treated more like our libraries and upgrade to hashicorp/go-metrics by utilizing its compat packages. This allows those importing the root module to control the metrics module used via build tags.
When we implemented CSI, the types of the fields for access mode and attachment
mode on volume requests were defined with a prefix "CSI". This gets confusing
now that we have dynamic host volumes using the same fields. Fortunately the
original was a typedef on string, and the Go API in the `api` package just uses
strings directly, so we can change the name of the type without breaking
backwards compatibility for the msgpack wire format.
Update the names to `VolumeAccessMode` and `VolumeAttachmentMode`. Keep the CSI
and DHV specific value constant names for these fields (they aren't currently
1:1), so that we can easily differentiate in a given bit of code which values
are valid.
Ref: https://github.com/hashicorp/nomad/pull/24881#discussion_r1920702890
The default node configuration in the client should always set an empty
HostVolumes map. Otherwise callers can panic, e.g.,:
goroutine 179 [running]:
github.com/hashicorp/nomad/client/hostvolumemanager.UpdateVolumeMap({0x36042b0, 0xc000c62a80}, 0x0, {0xc000a802a0, 0xd}, 0xc000691940)
github.com/hashicorp/nomad/client/hostvolumemanager/volume_fingerprint.go:43 +0x1b2
github.com/hashicorp/nomad/client.(*Client).batchFirstFingerprints.func1({0xc000a802a0, 0xd}, 0xc000691940)
github.com/hashicorp/nomad/client/node_updater.go:54 +0xd7
github.com/hashicorp/nomad/client.(*batchNodeUpdates).batchHostVolumeUpdates(0xc000912608?, 0xc0009f2f88)
github.com/hashicorp/nomad/client/node_updater.go:417 +0x152
github.com/hashicorp/nomad/client.(*Client).batchFirstFingerprints(0xc000c2d188)
github.com/hashicorp/nomad/client/node_updater.go:53 +0x1c5
created by github.com/hashicorp/nomad/client.NewClient in goroutine 1
github.com/hashicorp/nomad/client/client.go:557 +0x2069
is a panic of the HVM when restarting a client that doesn't have any static
host volumes, but does have a dynamic host volume.
The output of `GetDynamicHostVolumes` is a slice but that slice is constructed
from iterating over a map and isn't sorted. Sort the output in the test to
eliminate a test flake.
When we register a volume without a plugin, we need to send a client RPC so that
the node fingerprint can be updated. The registered volume also needs to be
written to client state so that we can restore the fingerprint after a restart.
Co-authored-by: Daniel Bennett <dbennett@hashicorp.com>
When the Nomad client restarts and restores allocations, the network namespace
for an allocation may exist but no longer be correctly configured. For example,
if the host is rebooted and the task was a Docker task using a pause container,
the network namespace may be recreated by the docker daemon.
When we restore an allocation, use the CNI "check" command to verify that any
existing network namespace matches the expected configuration. This requires CNI
plugins of at least version 1.2.0 to avoid a bug in older plugin versions that
would cause the check to fail.
If the check fails, destroy the network namespace and try to recreate it from
scratch once. If that fails in the second pass, fail the restore so that the
allocation can be recreated (rather than silently having networking fail).
This should fix the gap left #24650 for Docker task drivers and any other
drivers with the `MustInitiateNetwork` capability.
Fixes: https://github.com/hashicorp/nomad/issues/24292
Ref: https://github.com/hashicorp/nomad/pull/24650
a node can have only one volume with a given name.
the scheduler prevents duplicates, but can only
do so after the server knows about the volume.
this prevents multiple concurrent creates being
called faster than the fingerprint/heartbeat interval.
users may still modify an existing volume only
if they set the `id` in the volume spec and
re-issue `nomad volume create`
if a *static* vol is added to config with a name
already being used by a dynamic volume, the
dynamic takes precedence, but log a warning.
Instead of a plugin `version` subcommand that responds with a string
(established in #24497), respond to a `fingerprint` command with a data
structure that we may extend in the future (such as plugin capabilities,
like size constraint support?). In the immediate term, it's still just the
version: `{"version": "0.0.1"}`
In addition to leaving the door open for future expansion, I think it will
also avoid false positives detecting executables that just happen to
respond to a `version` command.
This also reverses the ordering of the fingerprint string parts
from `plugins.host_volume.version.mkdir` (which aligned with CNI)
to `plugins.host_volume.mkdir.version` (makes more sense to me)
