We are observing incorrect data in Prometheus, which is consistently
non-reproducible. After a restart, the issue does not occur, but if the
pods run for an extended period, they seem to enter a strange state
where the counters become entangled and start sharing arbitrary values
that are added to the counters.
For example, the `feature_lifecycle_stage_entered` counter gets an
arbitrary value, such as 12, added when `inc()` is called. The
`exceedsLimitErrorCounter` shows the same behavior, and the code
implementation is identical.
We also tested some existing `increase()` counters, and they do not
suffer from this issue.
All calls to `counter.labels(labels).inc(`) will be replaced by
`counter.increment()` to try to mitigate the issue.
Previously we expected the tag to look like `type:value`. Now we allow
everything after first colon, as the value and not break query
`type:this:still:is:value`.
Fixes a bug where if you had API keys using different casing for the
same type, they'd come out as different types in the API token count
map. To get around it, we normalize the keys to lowercase before
inserting them into the map, taking into account any previous values
that might have existed for that type.
Should fix issues like this:
Updates the instance stats endpoint with
- maxEnvironmentStrategies
- maxConstraints
- maxConstraintValues
It adds the following rows to the front end table:
- segments (already in the payload, just not used for the table before)
- API tokens (separate rows for type, + one for total) (also existed
before, but wasn't listed)
- Highest number of strategies used for a single flag in a single
environment
- Highest number of constraints used on a single strategy
- Highest number of values used for a single constraint
Turns out we've been trying to return API token data in instance stats
for a while, but that the serialization has failed. Serializing a JS map
just yields an empty object.
This PR fixes that serialization and also adds API tokens to the
instance stats schema (it wasn't before, but we did return it). Adding
it to the schema is also part of making resource usage visible as part
of the soft limits project.
Path types in our openapi are inferred as string (which is a sensible
default). But we can be more specific and provide the right type for
each parameter. This is one example of how we can do that
This PR fixes an issue where the number of flags belonging to a project
was wrong in the new getProjectsForAdminUi.
The cause was that we now join with the events table to get the most
"lastUpdatedAt" data. This meant that we got multiple rows for each
flag, so we counted the same flag multiple times. The fix was to use a
"distinct".
Additionally, I used this as an excuse to write some more tests that I'd
been thinking about. And in doing so also uncovered another bug that
would only ever surface in verrry rare conditions: if a flag had been
created in project A, but moved to project B AND the
feature-project-change event hadn't fired correctly, project B's last
updated could show data from that feature in project A.
I've also taken the liberty of doing a little bit of cleanup.
## About the changes
When storing last seen metrics we no longer validate at insert time that
the feature exists. Instead, there's a job cleaning up on a regular
interval.
Metrics for features with more than 255 characters, makes the whole
batch to fail, resulting in metrics being lost.
This PR helps mitigate the issue while also logs long name feature names
Implements empty responses for the fake project read model. Instead of
throwing a not implemented error, we'll return empty results.
This makes some of the tests in enterprise pass.
This PR touches up a few small things in the project read model.
Fixes:
Use the right method name in the query/method timer for
`getProjectsForAdminUi`. I'd forgotten to change the timer name from the
original method name.
Spells the method name correctly for the `getMembersCount` timer (it
used to be `getMemberCount`, but the method is callled `getMembersCount`
with a plural s).
Changes:
Call the `getMembersCount` timer from within the `getMembersCount`
method itself. Instead of setting that timer up from two different
places, we can call it in the method we're timing. This wasn't a problem
previously, because the method was only called from a single place.
Assuming we always wanna time that query, it makes more sense to put the
timing in the actual method.
Hooks up the new project read model and updates the existing project
service to use it instead when the flag is on.
In doing:
- creates a composition root for the read model
- includes it in IUnleashStores
- updates some existing methods to accept either the old or the new
model
- updates the OpenAPI schema to deprecate the old properties
These are both related to the work on the project list improvements
project.
The `projectListImprovements` flag will be used to enable disable the
new project list improvements.
The `useProjectReadModel` flag will be used to enable/disable the use
of the new project read model and is mostly a safety feature.
Creates a new project read model exposing data to be used for the UI and
for the insights module.
The model contains two public methods, both based on the project store's
`getProjectsWithCounts`:
- `getProjectsForAdminUi`
- `getProjectsForInsights`
This mirrors the two places where the base query is actually in use
today and adapts the query to those two explicit cases.
The new `getProjectsForAdminUi` method also contains data for last flag
update and last flag metric reported, as required for the new projects
list screen.
Additionally the read model contains a private `getMembersCount` method,
which is also lifted from the project store. This method was only used
in the old `getProjectsWithCounts` method, so I have also removed the
method from the public interface.
This PR does *not* hook up the new read model to anything or delete any
existing uses of the old method.
## Why?
As mentioned in the background, this query is used in two places, both
to get data for the UI (directly or indirectly). This is consistent with
the principles laid out in our [ADR on read vs write
models](https://docs.getunleash.io/contributing/ADRs/back-end/write-model-vs-read-models).
There is an argument to be made, however, that the insights module uses
this as an **internal** read model, but the description of an internal
model ("Internal read models are typically narrowly focused on answering
one question and usually require simple queries compared to external
read models") does not apply here. It's closer to the description of
external read models: "View model will typically join data across a few
DB tables" for display in the UI.
## Discussion points
### What about properties on the schema that are now gone?
The `project-schema`, which is delivered to the UI through the
`getProjects` endpoint (and nowhere else, it seems), describes
properties that will no longer be sent to the front end, including
`defaultStickiness`, `avgTimeToProduction`, and more. Can we just stop
sending them or is that a breaking change?
The schema does not define them as required properties, so in theory,
not sending them isn't breaking any contracts. We can deprecate the
properties and just not populate them anymore.
At least that's my thought on it. I'm open to hearing other views.
### Can we add the properties in fewer lines of code?
Yes! The [first commit in this PR
(b7534bfa)](b7534bfa07)
adds the two new properties in 8 lines of code.
However, this comes at the cost of diluting the `getProjectsWithCounts`
method further by adding more properties that are not used by the
insights module. That said, that might be a worthwhile tradeoff.
## Background
_(More [details in internal slack
thread](https://unleash-internal.slack.com/archives/C046LV6HH6W/p1723716675436829))_
I noticed that the project store's `getProjectWithCounts` is used in
exactly two places:
1. In the project service method which maps directly to the project
controller (in both OSS and enterprise).
2. In the insights service in enterprise.
In the case of the controller, that’s the termination point. I’d guess
that when written, the store only served the purpose of showing data to
the UI.
In the event of the insights service, the data is mapped in
getProjectStats.
But I was a little surprised that they were sharing the same query, so I
decided to dig a little deeper to see what we’re actually using and what
we’re not (including the potential new columns). Here’s what I found.
Of the 14 already existing properties, insights use only 7 and the
project list UI uses only 10 (though the schema mentions all 14 (as far
as I could tell from scouring the code base)). Additionally, there’s two
properties that I couldn’t find any evidence of being used by either:
- default stickiness
- updatedAt (this is when the project was last updated; not its flags)
