This commit replaces the use of "sanity" with more inclusive
naming.
When `sanity` is used in a more general sense either `validity`
or `quick` is used.
The snap client would always retry requests for a long time,
regardless of error type (e.g., snap list retries for 2min,
if it fails to read auth.json).
This commit prevents retries caused by internal and auth errors.
It is slightly unfortunate in that errors from the daemon that are not wrapped,
such as the unit test we have end up being generic, but most realistic errors
should already be quite readable and usable, so just remove the wrapping.
Signed-off-by: Ian Johnson <ian.johnson@canonical.com>
Instead of passing the values as strings, pass them as ints exactly with a map
of string -> interface{}. This induces a fair amount of overhead to parse
everywhere but still enables us to relatively easily add new keys/resource
types, assuming those types don't need special formatting like the memory does
with units.
Signed-off-by: Ian Johnson <ian.johnson@canonical.com>
Use a more proper error kind for quota conflicts, supporting the error type
*servicestate.QuotaChangeConflictError inside errToResponse and indicating what
sort of change the conflict is being used for.
Also update the unit tests for the daemon error response to check for this.
Signed-off-by: Ian Johnson <ian.johnson@canonical.com>
The CLI has organization where the current and constraint keys are maps of the
unit of measure for the quota group to the value, so update the REST API to
match this.
Signed-off-by: Ian Johnson <ian.johnson@canonical.com>
The exported methods from the servicestate package all return tasksets now,
which is expected to be put into changes that are executed by the overlord
loop. This involves changes in many parts that use quotas, such as the tests,
where a new mock function to create a quota group in state without running any
servicectl commands as well as changes to the client side.
Signed-off-by: Ian Johnson <ian.johnson@canonical.com>
had to make the comment in wrappers/services_test.go a one-liner
otherwise half of it is lost
last set of files needing changing (as per current master)
* client/quota.go: support current-memory key
Signed-off-by: Ian Johnson <ian.johnson@canonical.com>
* cmd/snap/quota: refactor quota CLI as per new design
* Make quota command only responsible for showing/displaying quota group
information.
* Introduce new set-quota command which creates or updates quota groups
* Move the display of memory limit under a map section "constraints" which then
has a memory key to allow future resource types to live under this section.
* Display current memory usage for quota groups under a new section "current",
which like "constraints" can be expanded for future resource types too.
Also update the spread test and unit tests for the new output formats.
Signed-off-by: Ian Johnson <ian.johnson@canonical.com>
* tests/main/snap-quota-groups: check the cgroup procs file in a loop
This too is a race condition, thanks to Maciej for pointing this out.
Signed-off-by: Ian Johnson <ian.johnson@canonical.com>
* tests/snap-quota-groups: compare slice mem usage from both snapd and kernel
This check ensures that what the kernel says the current memory usage is and
what snapd says the current memory usage is don't differ by more than 10%. In
practice they should be exactly equal if the program is doing nothing as the
go-example-webserver should be, but something does happen to change it \
shouldn't change drastically.
Signed-off-by: Ian Johnson <ian.johnson@canonical.com>
* tests/snap-quota: add a snap with a service to one of the groups
This ensures that we can check the `snap quota` and `snap quotas` output for
the "current" section with the memory usage in it.
Signed-off-by: Ian Johnson <ian.johnson@canonical.com>
* tests/main/snap-quota: fix trusty check
Signed-off-by: Ian Johnson <ian.johnson@canonical.com>
* tests/main/snap-quota: increase limits so the webserver snap can start
These limits were too low so the service was killed and no memory usage was
reported in the output, but we want to see some usage from the output to check
that it works, so increase the limits so the server is not killed due to OOM.
Signed-off-by: Ian Johnson <ian.johnson@canonical.com>
* tests/main/snap-quota: fix group-one line in output
Also fix the defer statement which was missing the snap to unset the config on.
Signed-off-by: Ian Johnson <ian.johnson@canonical.com>
* tests/main/snap-quota-groups: refactor memory check for empty quota groups
We only should be checking what snapd says about memory usage for these empty
groups, we don't need to compare with what the kernel says.
Signed-off-by: Ian Johnson <ian.johnson@canonical.com>
* tests/snap-quota-groups: fix typo
Signed-off-by: Ian Johnson <ian.johnson@canonical.com>
* tests/main/snap-quota-groups: treat memory=4096 as memory=0 for new systems
On the following systems:
- Arch Linux
- Fedora 33
- Fedora 34
- Debian sid
- Ubuntu 21.04
The memory usage of an "empty" but active cgroup ends up being 4K for some
reason, so handle this in the expected output.
Signed-off-by: Ian Johnson <ian.johnson@canonical.com>
* tests/snap-quota-groups: silly typos
Signed-off-by: Ian Johnson <ian.johnson@canonical.com>
* tests/main/snap-quota: fix more typos
non capturing groups are not a thing in grep -E because of course they aren't
Signed-off-by: Ian Johnson <ian.johnson@canonical.com>
* tests/main/snap-quota-groups: use python if python3 is not available
Signed-off-by: Ian Johnson <ian.johnson@canonical.com>
* tests/main/snap-quota: adjust memory limits for test groups to be at least 4K
On some systems, an empty cgroup with just cgroups nested inside it (but not
necessarily any processes) will have 4K memory usage, so on these systems, we
should make sure that the nested cgroups have enough space in them.
In a follow-up, we will adjust the minimum usable memory limit for a given
cgroup to be 4K to prevent this situation from happening in practice.
Signed-off-by: Ian Johnson <ian.johnson@canonical.com>
