On very small clusters, the node count heuristic is impractical and
leads to confusion. By additionally requiring 10+ sibling allocs, the
lines will be shown more often.
The connect tests are very disruptive: restart consul/nomad agents with new
tokens. The test seems particularly flaky, failing 32 times out of 73 in my
sample.
The tests are particularly problematic because they are disruptive and affect
other tests. On failure, the nomad or consul agent on the client can get into a
wedged state, so health/deployment info in subsequent tests may be wrong. In
some cases, the node will be deemed as fail, and then the subsequent tests may
fail when the node is deemed lost and the test allocations get migrated unexpectedly.
The nodedrain deadline test asserts that all allocations are migrated by the
deadline. However, when the deadline is short (e.g. 10s), the test may fail
because of scheduler/client-propagation delays.
In one failing test, it took ~15s from the RPC call to the moment to the moment
the scheduler issued migration update, and then 3 seconds for the alloc to be
stopped.
Here, I increase the timeouts to avoid such false positives.
Increase the timeout for vaultsecrets. As the default interval is 0.1s, 10
retries mean it only retries for one second, a very short time for some waiting
scenarios in the test (e.g. starting allocs, etc).
Prefer testutil.WaitForResultRetries that emits more descriptive errors on
failures. `require.Evatually` fails with opaque "Condition never satisfied"
error message.
This is an attempt at deflaking the e2e exec tests, and a way to improve
messages.
e2e occasionally fail with "unexpected EOF" even though the exec output matches
expectations. I suspect there is a race in handling EOF in server/http handling.
Here, we special case this error and ensures we get all failures,
to help debug the case better.
This fixes a couple bugs
1. Overreporting resources reserved due to counting terminal allocs
2. Overreporting unique client placements due to uniquing on object refs
instead of on client ID.
This PR makes two ergonomics changes, meant to get e2e builds more reproducible and ease changes.
### AMI Management
First, we pin the server AMIs to the commits associated with the build. No more using the latest AMI a developer build in a test branch, or accidentally using a stale AMI because we forgot to build one! Packer is to tag the AMI images with the commit sha used to generate the image, and then Terraform would look up only the AMIs associated with that sha. To minimize churn, we use the SHA associated with the latest Packer configurations, rather than SHA of all.
This has few benefits: reproducibility and avoiding accidental AMI changes and contamination of changes across branches. Also, the change is a stepping stone to an e2e pipeline that builds new AMIs automatically if Packer files changed.
The downside is that new AMIs will be generated even for irrelevant changes (e.g. spelling, commits), but I suspect that's OK. Also, an engineer will be forced to build the AMI whenever they change Packer files while iterating on e2e scripts; this hasn't been an issue for me yet, and I'll be open for iterating on that later if it proves to be an issue.
### Config Files and Packer
Second, this PR moves e2e config hcl management to Terraform instead of Packer. Currently, the config files live in `./terraform/config`, but they are baked into the servers by Packer and changes are ignored. This current behavior surprised me, as I spent a bit of time debugging why my config changes weren't applied. Having Terraform manage them would ease engineer's iteration. Also, make Packer management more consistent (Packer only works `e2e/terraform/packer`), and easing the logic for AMI change detection.
The config directory is very small (100KB), and having it as an upload step adds negligible time to `terraform apply`.
