mirror of
https://github.com/cozystack/cozystack.git
synced 2026-10-26 21:17:32 +09:00
Myasnikov Daniil
cc942fed68
The shell translator rewrites test bodies and loses Bats reporting. The unit lane needs complete TAP results and JUnit reports without changing how live-cluster suites run. Sourced production helpers still need a dash pass, and unit test bodies need nounset without exporting shell options into subprocesses. Assisted-by: LLM Signed-off-by: Myasnikov Daniil <myasnikovdaniil2001@gmail.com>
95 lines
4.5 KiB
Bash
95 lines
4.5 KiB
Bash
#!/usr/bin/env bats
|
|
# -----------------------------------------------------------------------------
|
|
# Unit tests for the probe-output resolver in
|
|
# hack/e2e-chainsaw/_lib/etcd-probe.sh.
|
|
#
|
|
# The etcd suite proves the member Pods serve Prometheus metrics over plaintext
|
|
# http by running a throwaway curl Pod and reading back a `code=<status>` marker.
|
|
# `kubectl run --attach` is how that marker comes back, and it loses the stream
|
|
# of a container that exits before the attach connection is established -- curl
|
|
# against a Pod IP finishes in milliseconds, so the loss is routine rather than
|
|
# exotic. When it happens the
|
|
# caller sees only kubectl's own chatter ("If you don't see a command prompt,
|
|
# try pressing enter."), reads no marker, and reports the metrics endpoint as
|
|
# broken -- a false negative on a healthy cluster.
|
|
#
|
|
# etcd_probe_resolve is the recovery: the container's stdout is still on disk in
|
|
# the Pod's log, so a lost attach stream is re-read rather than believed. These
|
|
# tests pin that recovery, the pass-through when attach did carry the marker,
|
|
# and the honest surfacing of a genuine non-200 (which must NOT be masked by the
|
|
# fallback -- a real failure has to stay a failure).
|
|
#
|
|
# The resolver is separated from the kubectl orchestration precisely so it can
|
|
# be tested without a cluster; the `kubectl run`/`kubectl logs` calls around it
|
|
# need a live cluster and are covered by the e2e run, matching how the other
|
|
# hack/ helpers are tested.
|
|
#
|
|
# CI runs this file under Bats through `make bats-unit-tests`. It also remains
|
|
# compatible with the legacy `hack/cozytest.sh` translator, whose awk parser
|
|
# recognizes only @test blocks and a bare `}` on its own line. The tests avoid
|
|
# `run`, `$status`, setup(), and teardown() for that compatibility path and use
|
|
# direct shell assertions. Titles remain distinguishable after stripping all
|
|
# but [A-Za-z0-9], as required by the legacy translator.
|
|
#
|
|
# Run with: bats hack/etcd-probe_test.bats
|
|
# -----------------------------------------------------------------------------
|
|
|
|
load test_helper
|
|
|
|
HACK_DIR="$(cd "$(dirname "${BATS_TEST_FILENAME:-$0}")" && pwd)"
|
|
# shellcheck source=/dev/null
|
|
. "$HACK_DIR/e2e-chainsaw/_lib/etcd-probe.sh"
|
|
|
|
@test "resolver returns the attach output unchanged when it carries the marker" {
|
|
out="$(etcd_probe_resolve 'code=200' printf 'code=500')"
|
|
# The fallback must not run at all when attach already delivered a marker:
|
|
# re-reading the log would be a second source of truth for the same probe.
|
|
[ "$out" = "code=200" ]
|
|
}
|
|
|
|
@test "resolver rereads the pod log when the attach stream carried no marker" {
|
|
lost="If you don't see a command prompt, try pressing enter."
|
|
out="$(etcd_probe_resolve "$lost" printf 'code=200')"
|
|
# This is the regression under test: before the fallback existed the lost
|
|
# stream was reported verbatim as the probe result and failed the suite on a
|
|
# cluster whose metrics endpoint was serving 200 the whole time.
|
|
[ "$out" = "code=200" ]
|
|
}
|
|
|
|
@test "resolver rereads the pod log when the attach stream was entirely empty" {
|
|
out="$(etcd_probe_resolve '' printf 'code=200')"
|
|
[ "$out" = "code=200" ]
|
|
}
|
|
|
|
@test "resolver surfaces a genuine non200 from the attach stream" {
|
|
# A real failure must survive the resolver untouched. If this ever returns
|
|
# the fallback's value instead, the fallback has become a way to launder a
|
|
# broken metrics endpoint into a pass.
|
|
out="$(etcd_probe_resolve 'code=503' printf 'code=200')"
|
|
[ "$out" = "code=503" ]
|
|
}
|
|
|
|
@test "resolver surfaces a genuine non200 recovered from the pod log" {
|
|
lost="If you don't see a command prompt, try pressing enter."
|
|
out="$(etcd_probe_resolve "$lost" printf 'code=503')"
|
|
[ "$out" = "code=503" ]
|
|
}
|
|
|
|
@test "resolver keeps the lost attach output when the log reread also yields nothing" {
|
|
# Both sources empty means the probe genuinely produced no marker. Returning
|
|
# the attach text keeps whatever kubectl said about why in the failure
|
|
# message, instead of reporting an empty string the reader cannot act on.
|
|
lost="Error from server: pods etcd-metrics-probe not found"
|
|
out="$(etcd_probe_resolve "$lost" printf '')"
|
|
[ "$out" = "$lost" ]
|
|
}
|
|
|
|
@test "resolver tolerates a failing log reread without aborting the caller" {
|
|
# `kubectl logs` against an evicted or already-collected Pod exits non-zero.
|
|
# Under the suite's `set -e` that must not kill the step before it can
|
|
# report the probe result.
|
|
lost="If you don't see a command prompt, try pressing enter."
|
|
out="$(etcd_probe_resolve "$lost" false)"
|
|
[ "$out" = "$lost" ]
|
|
}
|