When the scheduler tries to find a placement for a new allocation, it iterates
over a subset of nodes. For each node, we populate a `NetworkIndex` bitmap with
the ports of all existing allocations and any other allocations already proposed
as part of this same evaluation via its `SetAllocs` method. Then we make an
"ask" of the `NetworkIndex` in `AssignPorts` for any ports we need and receive
an "offer" in return. The offer will include both static ports and any dynamic
port assignments.
The `AssignPorts` method was written to support group networks, and it shares
code that selects dynamic ports with the original `AssignTaskNetwork`
code. `AssignTaskNetwork` can request multiple ports from the bitmap at a
time. But `AssignPorts` requests them one at a time and does not account for
possible collisions, and doesn't return an error in that case.
What happens next varies:
1. If the scheduler doesn't place the allocation on that node, the port
conflict is thrown away and there's no problem.
2. If the node is picked and this is the only allocation (or last allocation),
the plan applier will reject the plan when it calls `SetAllocs`, as we'd expect.
3. If the node is picked and there are additional allocations in the same eval
that iterate over the same node, their call to `SetAllocs` will detect the
impossible state and the node will be rejected. This can have the puzzling
behavior where a second task group for the job without any networking at all
can hit a port collision error!
It looks like this bug has existed since we implemented group networks, but
there are several factors that add up to making the issue rare for many users
yet frustratingly frequent for others:
* You're more likely to hit this bug the more tightly packed your range for
dynamic ports is. With 12000 ports in the range by default, many clusters can
avoid this for a long time.
* You're more likely to hit case (3) for jobs with lots of allocations or if a
scheduler has to iterate over a large number of nodes, such as with system jobs,
jobs with `spread` blocks, or (sometimes) jobs using `unique` constraints.
For unlucky combinations of these factors, it's possible that case (3) happens
repeatedly, preventing scheduling of a given job until a client state
change (ex. restarting the agent so all its allocations are rescheduled
elsewhere) re-opens the range of dynamic ports available.
This changeset:
* Fixes the bug by accounting for collisions in dynamic port selection in
`AssignPorts`.
* Adds test coverage for `AssignPorts`, expands coverage of this case for the
deprecated `AssignTaskNetwork`, and tightens the dynamic port range in a
scheduler test for spread scheduling to more easily detect this kind of problem
in the future.
* Adds a `String()` method to `Bitmap` so that any future "screaming" log lines
have a human-readable list of used ports.
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.
When a node becomes ready, create an eval for all system jobs across
namespaces.
The previous code uses `job.ID` to deduplicate evals, but that ignores
the job namespace. Thus if there are multiple jobs in different
namespaces sharing the same ID/Name, only one will be considered for
running in the new node. Thus, Nomad may skip running some system jobs
in that node.
This PR implements a new "System Batch" scheduler type. Jobs can
make use of this new scheduler by setting their type to 'sysbatch'.
Like the name implies, sysbatch can be thought of as a hybrid between
system and batch jobs - it is for running short lived jobs intended to
run on every compatible node in the cluster.
As with batch jobs, sysbatch jobs can also be periodic and/or parameterized
dispatch jobs. A sysbatch job is considered complete when it has been run
on all compatible nodes until reaching a terminal state (success or failed
on retries).
Feasibility and preemption are governed the same as with system jobs. In
this PR, the update stanza is not yet supported. The update stanza is sill
limited in functionality for the underlying system scheduler, and is
not useful yet for sysbatch jobs. Further work in #4740 will improve
support for the update stanza and deployments.
Closes#2527
Cluster operators want to have better control over memory
oversubscription and may want to enable/disable it based on their
experience.
This PR adds a scheduler configuration field to control memory
oversubscription. It's additional field that can be set in the [API via Scheduler Config](https://www.nomadproject.io/api-docs/operator/scheduler), or [the agent server config](https://www.nomadproject.io/docs/configuration/server#configuring-scheduler-config).
I opted to have the memory oversubscription be an opt-in, but happy to change it. To enable it, operators should call the API with:
```json
{
"MemoryOversubscriptionEnabled": true
}
```
If memory oversubscription is disabled, submitting jobs specifying `memory_max` will get a "Memory oversubscription is not
enabled" warnings, but the jobs will be accepted without them accessing
the additional memory.
The warning message is like:
```
$ nomad job run /tmp/j
Job Warnings:
1 warning(s):
* Memory oversubscription is not enabled; Task cache.redis memory_max value will be ignored
==> Monitoring evaluation "7c444157"
Evaluation triggered by job "example"
==> Monitoring evaluation "7c444157"
Evaluation within deployment: "9d826f13"
Allocation "aa5c3cad" created: node "9272088e", group "cache"
Evaluation status changed: "pending" -> "complete"
==> Evaluation "7c444157" finished with status "complete"
# then you can examine the Alloc AllocatedResources to validate whether the task is allowed to exceed memory:
$ nomad alloc status -json aa5c3cad | jq '.AllocatedResources.Tasks["redis"].Memory'
{
"MemoryMB": 256,
"MemoryMaxMB": 0
}
```
Start tracking a new MemoryMaxMB field that represents the maximum memory a task
may use in the client. This allows tasks to specify a memory reservation (to be
used by scheduler when placing the task) but use excess memory used on the
client if the client has any.
This commit adds the server tracking for the value, and ensures that allocations
AllocatedResource fields include the value.
* ar: support opting into binding host ports to default network IP
* fix config plumbing
* plumb node address into network resource
* struct: only handle network resource upgrade path once
We currently log an error if preemption is unable to find a suitable set of
allocations to preempt. This commit changes that to debug level since not finding
preemptable allocations is not an error condition.
This adds constraints for asserting that a given attribute or value
exists, or does not exist. This acts as a companion to =, or !=
operators, e.g:
```hcl
constraint {
attribute = "${attrs.type}"
operator = "!="
value = "database"
}
constraint {
attribute = "${attrs.type}"
operator = "is_set"
}
```
Also changes the logic for score when there is more than one task
requesting a device. Since inter task affinities are already normalized,
we take the average of the scores across tasks.
This commit implements an allocation selection algorithm for finding
allocations to preempt. It currently special cases network resource asks
from others (cpu/memory/disk/iops).
