When running at scale, it's possible that Docker Engine starts
containers successfully but gets wedged in a way where API call fails.
The Docker Engine may remain unavailable for arbitrary long time.
Here, we introduce a periodic reconcilation process that ensures that any
container started by nomad is tracked, and killed if is running
unexpectedly.
Basically, the periodic job inspects any container that isn't tracked in
its handlers. A creation grace period is used to prevent killing newly
created containers that aren't registered yet.
Also, we aim to avoid killing unrelated containters started by host or
through raw_exec drivers. The logic is to pattern against containers
environment variables and mounts to infer if they are an alloc docker
container.
Lastly, the periodic job can be disabled to avoid any interference if
need be.
This commit introduces support for configuring mount propagation when
mounting volumes with the `volume_mount` stanza on Linux targets.
Similar to Kubernetes, we expose 3 options for configuring mount
propagation:
- private, which is equivalent to `rprivate` on Linux, which does not allow the
container to see any new nested mounts after the chroot was created.
- host-to-task, which is equivalent to `rslave` on Linux, which allows new mounts
that have been created _outside of the container_ to be visible
inside the container after the chroot is created.
- bidirectional, which is equivalent to `rshared` on Linux, which allows both
the container to see new mounts created on the host, but
importantly _allows the container to create mounts that are
visible in other containers an don the host_
private and host-to-task are safe, but bidirectional mounts can be
dangerous, as if the code inside a container creates a mount, and does
not clean it up before tearing down the container, it can cause bad
things to happen inside the kernel.
To add a layer of safety here, we require that the user has ReadWrite
permissions on the volume before allowing bidirectional mounts, as a
defense in depth / validation case, although creating mounts should also require
a priviliged execution environment inside the container.
The docker creation API calls may fail with http errors (e.g. timeout)
even if container was successfully created.
Here, we force remove container if we got unexpected failure. We
already do this in some error handlers, and this commit updates all
paths.
I stopped short from a more aggressive refactoring, as the code is ripe
for refactoring and would rather do that in another PR.
This handles a bug where we may start a container successfully, yet we
fail due to retries and startContainer not being idempotent call.
Here, we ensure that when starting a container fails with 500 error,
the retry succeeds if container was started successfully.
hclspec.NewLiteral does not quote its values, which caused `3m` to be
parsed as a nonsensical literal which broke the plugin loader during
initialization. By quoting the value here, it starts correctly.
Using the API as provided from the `mounts` package imposes validation
on the `src:dest` which shouldn't be performed at this time. To workaround
that lift the internal code from that library required to only perform
the split.
Currently, nomad "plugin" processes (e.g. executor, logmon, docker_logger) are started as CLI
commands to be handled by command CLI framework. Plugin launchers use
`discover.NomadBinary()` to identify the binary and start it.
This has few downsides: The trivial one is that when running tests, one
must re-compile the nomad binary as the tests need to invoke the nomad
executable to start plugin. This is frequently overlooked, resulting in
puzzlement.
The more significant issue with `executor` in particular is in relation
to external driver:
* Plugin must identify the path of invoking nomad binary, which is not
trivial; `discvoer.NomadBinary()` now returns the path to the plugin
rather than to nomad, preventing external drivers from launching
executors.
* The external driver may get a different version of executor than it
expects (specially if we make a binary incompatible change in future).
This commit addresses both downside by having the plugin invocation
handling through an `init()` call, similar to how libcontainer init
handler is done in [1] and recommened by libcontainer [2]. `init()`
will be invoked and handled properly in tests and external drivers.
For external drivers, this change will cause external drivers to launch
the executor that's compiled against.
There a are a couple of downsides to this approach:
* These specific packages (i.e executor, logmon, and dockerlog) need to
be careful in use of `init()`, package initializers. Must avoid having
command execution rely on any other init in the package. I prefixed
files with `z_` (golang processes files in lexical order), but ensured
we don't depend on order.
* The command handling is spread in multiple packages making it a bit
less obvious how plugin starts are handled.
[1] drivers/shared/executor/libcontainer_nsenter_linux.go
[2] eb4aeed24f/libcontainer (using-libcontainer)
Support Docker `volumes` field in Windows. Previously, volumes parser
assumed some Unix-ism (e.g. didn't expect `:` in mount paths).
Here, we use the Docker parser to identify host and container paths.
Docker parsers use different validation logic from our previous unix
implementation: Docker parser accepts single path as a volume entry
(parsing it as a container path with auto-created volume) and enforces
additional checks (e.g. validity of mode). Thereforce, I opted to use
Docker parser only for Windows, and keep Nomad's linux parser to
preserve current behavior.
In Nomad 0.9, we made volume driver handling the same for `""`, and
`"local"` volumes. Prior to Nomad 0.9 however these had slightly different
behaviour for relative paths and named volumes.
Prior to 0.9 the empty string would expand relative paths within the task
dir, and `"local"` volumes that are not absolute paths would be treated
as docker named volumes.
This commit reverts to the previous behaviour as follows:
| Nomad Version | Driver | Volume Spec | Behaviour |
|-------------------------------------------------------------------------
| all | "" | testing:/testing | allocdir/testing |
| 0.8.7 | "local" | testing:/testing | "testing" as named volume |
| 0.9.0 | "local" | testing:/testing | allocdir/testing |
| 0.9.1 | "local" | testing:/testing | "testing" as named volume |
nvidia library use of dynamic library seems to conflict with alpine and
musl based OSes. This adds a `nonvidia` tag to allow compiling nomad
for alpine images.
The nomad releases currently only support glibc based OS environments,
so we default to compiling with nvidia.
Fix AppVeyor failing builds, by moving docker image url test to run on unix
systems only. The used paused image is a linux image only, not
available on Windows.
destCh was being written to by one goroutine and closed by another
goroutine. This panic occurred in Travis:
```
=== FAIL: drivers/docker TestDockerCoordinator_ConcurrentPulls (117.66s)
=== PAUSE TestDockerCoordinator_ConcurrentPulls
=== CONT TestDockerCoordinator_ConcurrentPulls
panic: send on closed channel
goroutine 5358 [running]:
github.com/hashicorp/nomad/drivers/docker.dockerStatsCollector(0xc0003a4a20, 0xc0003a49c0, 0x3b9aca00)
/home/travis/gopath/src/github.com/hashicorp/nomad/drivers/docker/stats.go:108 +0x167
created by
github.com/hashicorp/nomad/drivers/docker.TestDriver_DockerStatsCollector
/home/travis/gopath/src/github.com/hashicorp/nomad/drivers/docker/stats_test.go:33 +0x1ab
```
The 2 ways to fix this kind of error are to either (1) add extra
coordination around multiple goroutines writing to a chan or (2) make it
so only one goroutines writes to a chan.
I implemented (2) first as it's simpler, but @notnoop pointed out since
the same destCh in reused in the stats loop there's now a double close
panic possible!
So this implements (1) by adding a *usageSender struct for handling
concurrent senders and closing.
Fix#5418
When using a docker logger that doesn't support log streaming through
API, currently docker logger runs a tight loop of Docker API calls
unexpectedly. This change ensures we stop fetching logs early.
Also, this adds some basic backoff strategy when Docker API logging
fails unexpectedly, to avoid accidentally DoSing the docker daemon.
Noticed that the protobuf files are out of sync with ones generated by 1.2.0 protoc go plugin.
The cause for these files seem to be related to release processes, e.g. [0.9.0-beta1 preperation](ecec3d38de (diff-da4da188ee496377d456025c2eab4e87)), and [0.9.0-beta3 preperation](b849d84f2f).
This restores the changes to that of the pinned protoc version and fails build if protobuf files are out of sync. Sample failing Travis job is that of the first commit change: https://travis-ci.org/hashicorp/nomad/jobs/506285085
This commit causes the docker driver to return undetected before it
first establishes a connection to the docker daemon.
This fixes a bug where hosts without docker installed would return as
unhealthy, rather than undetected.
Currently if a docker_logger cannot be reattached to, we will leak the
container that was being used. This is problematic if e.g using static
ports as it means you can never recover your task, or if a service is
expensive to run and will then be running without supervision.
Sometimes the nomad docker_logger may be killed by a service manager
when restarting the client for upgrades or reliability reasons.
Currently if this happens, we leak the users container and try to
reschedule over it.
This commit adds a new step to the recovery process that will spawn a
new docker logger process that will fetch logs from _the current
timestamp_. This is to avoid restarting users tasks because our logging
sidecar has failed.
This commit adds some extra resiliency to the docker logger in the case
of API failure from the docker daemon, by restarting the stream from the
current point in time if the stream returns and the container is still
running.
This ensures that `port_map` along with other block like attribute
declarations (e.g. ulimit, labels, etc) can handle various hcl and json
syntax that was supported in 0.8.
In 0.8.7, the following declarations are effectively equivalent:
```
// hcl block
port_map {
http = 80
https = 443
}
// hcl assignment
port_map = {
http = 80
https = 443
}
// json single element array of map (default in API response)
{"port_map": [{"http": 80, "https": 443}]}
// json array of individual maps (supported accidentally iiuc)
{"port_map: [{"http": 80}, {"https": 443}]}
```
We achieve compatbility by using `NewAttr("...", "list(map(string))",
false)` to be serialized to a `map[string]string` wrapper, instead of using
`BlockAttrs` declaration. The wrapper merges the list of maps
automatically, to ease driver development.
This approach is closer to how v0.8.7 implemented the fields [1][2], and
despite its verbosity, seems to perserve 0.8.7 behavior in hcl2.
This is only required for built-in types that have backward
compatibility constraints. External drivers should use `BlockAttrs`
instead, as they see fit.
[1] https://github.com/hashicorp/nomad/blob/v0.8.7/client/driver/docker.go#L216
[2] https://github.com/hashicorp/nomad/blob/v0.8.7/client/driver/docker.go#L698-L700
Windows Docker daemon does not support SIGINT, SIGTERM is the semantic
equivalent that allows for graceful shutdown before being followed up by
a SIGKILL.
* master: (23 commits)
tests: avoid assertion in goroutine
spell check
ci: run checkscripts
tests: deflake TestRktDriver_StartWaitRecoverWaitStop
drivers/rkt: Remove unused github.com/rkt/rkt
drivers/rkt: allow development on non-linux
cli: Hide `nomad docker_logger` from help output
api: test api and structs are in sync
goimports until make check is happy
nil check node resources to prevent panic
tr: use context in as select statement
move pluginutils -> helper/pluginutils
vet
goimports
gofmt
Split hclspec
move hclutils
Driver tests do not use hcl2/hcl, hclspec, or hclutils
move reattach config
loader and singleton
...
Due to https://github.com/tsenart/deadcode/issues/3 we can't specify
these consts on their own. This moves them into the _platform_test.go
files to avoid creating a package that only exposes a couple of values.
* Docker for Windows does not support ulimits
* Use filepath.ToSlash to test workdir
* Convert expected mount paths to system style
* Skip security-opt test on windows
- Windows does not support seccomp, and it's unclear which options are
available.
* Skip StartN due to lack of sigint
* docker: Use api to get image info on windows
* No bridge on windows
* Stop hardcoding /bin/
Uses the home directory and windows path expansion, as c:\tmp doesn't
necessarily exist, and mktemp would involve unnecessarily complicating
the commands.
- docker fingerprint issues a docker api system info call to get the
list of supported OCI runtimes.
- OCI runtimes are reported as comma separated list of names
- docker driver is aware of GPU runtime presence
- docker driver throws an error when user tries to run container with
GPU, when GPU runtime is not present
- docker GPU runtime name is configurable
Previously, we did not attempt to stop Docker Logger processes until
DestroyTask, which means that under many circumstances, we will never
successfully close the plugin client.
This commit terminates the plugin process when `run` terminates, or when
`DestroyTask` is called.
Steps to repro:
```
$ nomad agent -dev
$ nomad init
$ nomad run example.nomad
$ nomad stop example
$ ps aux | grep nomad # See docker logger process running
$ signal the dev agent
$ ps aux | grep nomad # See docker logger process running
```
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 fixes various instances of plugins being launched without using
the parent loggers. This meant that logs would not all go to the same
output, break formatting etc.
Use `/tmp` as temporary directory for docker driver tests, so tests can
run out of the box without any intervention.
macOS sets tempdir as `/var`, which Docker does not whitelist as a path
that can be bind-mounted.
Re-export the ResourceUsage structs in drivers package to avoid drivers
directly depending on the internal client/structs package directly.
I attempted moving the structs to drivers, but that caused some import
cycles that was a bit hard to disentagle. Alternatively, I added an
alias here that's sufficient for our purposes of avoiding external
drivers depend on internal packages, while allowing us to restructure
packages in future without breaking source compatibility.
This implements the InternalPluginDriver interface in each driver, and
calls the cancellation fn for their respective eventers.
This fixes a per task goroutine leak during test suite execution.
Currently the docker driver does not remove tasks from its state map
when destroying the task, which leads to issues when restarting tasks in
place, and leaks expired handles over time.
We already have two other Kill tests (e.g.
TestDockerDriver_Start_Kill_Wait and
TestDockerDriver_Start_KillTimeout), so don't need yet another flaky
test.
Noticed an issue in Docker daemon failing to handle the OOM test case
failure in build https://travis-ci.org/hashicorp/nomad/jobs/468027848 ,
and I suspect it's related to the process dying so quickly, and
potentially the way we are starting the task, so added a start up delay
and made it more consistent with other tests that don't seem as flaky.
The following is the log line showing Docker returning 500 error condition; while we can probably handle it gracefully without retrying, the retry is very cheap in this case and it's more of an optimization that we can handle in follow up PR.
```
testlog.go:32: 2018-12-14T14:57:52.626Z [DEBUG] docker/driver.go:852: docker: setting container startup command: task_name=nc-demo command="/bin/nc -l 127.0.0.1 -p 0"
testlog.go:32: 2018-12-14T14:57:52.626Z [DEBUG] docker/driver.go:866: docker: setting container name: task_name=nc-demo container_name=724a3e77-8b15-e657-f6aa-84c2d3243b18
testlog.go:32: 2018-12-14T14:57:52.694Z [INFO ] docker/driver.go:196: docker: created container: container_id=362b6ea183f3c4ce472d7d7571ca47023cea1df0f5eb920827921716f17718be
testlog.go:32: 2018-12-14T14:57:53.523Z [DEBUG] docker/driver.go:416: docker: failed to start container: container_id=362b6ea183f3c4ce472d7d7571ca47023cea1df0f5eb920827921716f17718be attempt=1 error="API error (500): {"message":"cannot start a stopped process: unknown"}
"
testlog.go:32: 2018-12-14T14:57:55.394Z [DEBUG] docker/driver.go:416: docker: failed to start container: container_id=362b6ea183f3c4ce472d7d7571ca47023cea1df0f5eb920827921716f17718be attempt=2 error="API error (500): {"message":"cannot start a stopped process: unknown"}
"
testlog.go:32: 2018-12-14T14:57:57.243Z [DEBUG] docker/driver.go:416: docker: failed to start container: container_id=362b6ea183f3c4ce472d7d7571ca47023cea1df0f5eb920827921716f17718be attempt=3 error="API error (500): {"message":"cannot start a stopped process: unknown"}
"
```
Using `:latest` tag is typically a cause of pain, as underlying image
changes behavior. Here, I'm switching to using a point release, and
re-updating the stored tarballs with it.
Sadly, when saving/loading images, the repo digeset is not supported:
https://github.com/moby/moby/issues/22011 ; but using point releases
should mitigate the problem.
The motivation here is that docker tests have some flakiness due to
accidental importing of `busybox:latest` which has `/bin/nc` that no
longer supports `-p 0`:
```
$ docker run -it --rm busybox /bin/nc -l 127.0.0.1 -p 0
Unable to find image 'busybox:latest' locally
latest: Pulling from library/busybox
Digest: sha256:2a03a6059f21e150ae84b0973863609494aad70f0a80eaeb64bddd8d92465812
Status: Downloaded newer image for busybox:latest
nc: bad local port '0'
```
Looks like older busybox versions (e.g. `busybox:1.24` do honor `-p 0`
as the test expect, but I would rather update busybox to fix.
Some tests have containers that die almost immediately, and may die
and cleaned up before `driver.WaitUntilStarted` runs.
The causes for container dying seems special for each test:
* TestDockerDriver_Cleanup: `hello-world` image just emits a message and exits immediately
* TestDockerDriver_ForcePull_RepoDigest: the busybox image in `TestDockerDriver_ForcePull_RepoDigest` test didn't support `-p 0` argument
* TestDockerDriver_Entrypoint: with the entrypoint being `/bin/sh -c`, the command needs to be the entire string; otherwise, it ignores the comments
this allows us to drop a cyclical import, but is subobptimal as it
requires BaseDriver tests to move. This falls firmly into the realm of
being a hack. Alternatives welcome.
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.
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.
This allows the container to be tagged with a user friendly image name
(e.g. `redis:3.2`) rather than the image ID (e.g.
`sha256:87856cc39862cec77541d68382e4867d7ccb29a85a17221446c857ddaebca916`).
Useful for human debugging, as well as some debugging and image scanning
tools.
This risks two bad changes:
1. Discrepancy in image resolution between docker and Nomad's image
loader.
* I checked the image creation paths in Nomad, and noticed that we
either pulled the image or inspect the image with the user provided
name.
2. A race in image tagging where the tag is modified between image
loading and container creation.
* I, personally, don't think this case is cause for concern, as it is
analogous to the task running a bit later. As long as the image is
still present, creating the container should be good.
