feat(P09): collector + aggregator (C-11/C-12/C-14) + drift aggregation (REQ-107)
scripts/orca-aggregate.sh: 10s aggregator, cluster.json merge + drift-events rsync + remediation trigger (P10b stub). scripts/orca- watchdog.sh: C-11 starvation detection. internal/cli/collector.go: orca collector start/stop/status. Tests: CLI + bats. ---ci--- project: orca phase: 09 milestone: v0.11 status: execute ---/ci---
This commit is contained in:
Jon Chery
Executable
+172
@@ -0,0 +1,172 @@
|
||||
#!/usr/bin/env bash
|
||||
# orca-aggregate.sh — lead-side aggregator (P09, C-11/C-12/C-14, REQ-107, D-237).
|
||||
#
|
||||
# Runs every 10s via orca-aggregate.timer. For each peer in
|
||||
# /etc/orca/peers.env: SSHs in, reads /run/orca/state/<latest>.json (the
|
||||
# peer's state snapshot), and merges all peer states into a single
|
||||
# /etc/orca/state/cluster.json on the lead. Writes atomically (temp +
|
||||
# rename). Structured logging via logger -t orca-aggregate (syslog).
|
||||
#
|
||||
# Drift-event aggregation extension (REQ-107, D-237): rsyncs each
|
||||
# peer's /etc/orca/state/drift-events/ to a temp dir, validates each
|
||||
# event's hash against /etc/orca/state/applied/<txn>/manifest.json,
|
||||
# triggers orca-remediate.sh <peer> <txn> for auto-remediable paths
|
||||
# (the remediation script itself lands in P10b — the call is emitted
|
||||
# but the script may not exist yet), then consumes (deletes) the event
|
||||
# file on the peer after processing. Aggregated drift state is written
|
||||
# to /etc/orca/state/drift-events-aggregated.json. The drift section
|
||||
# is skipped gracefully (with a log line) when the drift-events
|
||||
# directory does not exist on the lead or on any peer.
|
||||
#
|
||||
# The aggregator is opt-in (C-12): the operator enables it with
|
||||
# `orca cluster config --aggregator=true`, which installs this script
|
||||
# + the systemd units via `orca collector start`.
|
||||
|
||||
set -euo pipefail
|
||||
|
||||
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
|
||||
# shellcheck source=lib/orca-log.sh
|
||||
. "$SCRIPT_DIR/lib/orca-log.sh"
|
||||
|
||||
ORCA_LOG_ACTOR="spiffe://orca/cli/aggregator"
|
||||
|
||||
PEERS_ENV="${ORCA_PEERS_ENV:-/etc/orca/peers.env}"
|
||||
STATE_DIR="${ORCA_STATE_DIR:-/etc/orca/state}"
|
||||
CLUSTER_JSON="$STATE_DIR/cluster.json"
|
||||
DRIFT_AGG_JSON="$STATE_DIR/drift-events-aggregated.json"
|
||||
REMOTE_STATE_DIR="${ORCA_REMOTE_STATE_DIR:-/run/orca/state}"
|
||||
REMOTE_DRIFT_DIR="${ORCA_REMOTE_DRIFT_DIR:-/etc/orca/state/drift-events}"
|
||||
APPLIED_DIR="${ORCA_APPLIED_DIR:-/etc/orca/state/applied}"
|
||||
REMEDIATE_SCRIPT="${ORCA_REMEDIATE_SCRIPT:-/usr/local/sbin/orca-remediate.sh}"
|
||||
SSH_OPTS="${ORCA_SSH_OPTS:--o BatchMode=yes -o StrictHostKeyChecking=accept-new -o ConnectTimeout=5}"
|
||||
RSYNC_OPTS_RSYNC="${ORCA_RSYNC_OPTS_RSYNC:-}"
|
||||
AGG_TMP="$(mktemp -d)"
|
||||
trap 'rm -rf "$AGG_TMP"' EXIT
|
||||
|
||||
mkdir -p "$STATE_DIR"
|
||||
|
||||
read_peers() {
|
||||
if [ ! -f "$PEERS_ENV" ]; then
|
||||
return 0
|
||||
fi
|
||||
awk -F= '/^[[:space:]]*#/ {next} /^[[:space:]]*$/ {next} {sub(/^[[:space:]]*PEER_[^=]*=/,""); print}' "$PEERS_ENV"
|
||||
}
|
||||
|
||||
merge_tmp="$AGG_TMP/merge.jsonl"
|
||||
: >"$merge_tmp"
|
||||
|
||||
peer_count=0
|
||||
while IFS= read -r peer; do
|
||||
[ -z "$peer" ] && continue
|
||||
peer_count=$((peer_count + 1))
|
||||
latest_json="$(ssh $SSH_OPTS "$peer" "ls -1 $REMOTE_STATE_DIR/*.json 2>/dev/null | sort | tail -1" 2>/dev/null || true)"
|
||||
if [ -z "$latest_json" ]; then
|
||||
orca_log_warn "aggregate" "$peer" "skipped" "no state snapshot on peer"
|
||||
continue
|
||||
fi
|
||||
snapshot="$(ssh $SSH_OPTS "$peer" "cat '$latest_json'" 2>/dev/null || true)"
|
||||
if [ -z "$snapshot" ]; then
|
||||
orca_log_warn "aggregate" "$peer" "skipped" "failed to read state snapshot"
|
||||
continue
|
||||
fi
|
||||
printf '{"peer":"%s","state":%s}\n' "$peer" "$snapshot" >>"$merge_tmp"
|
||||
orca_log_info "aggregate" "$peer" "ok" "snapshot=$latest_json"
|
||||
done < <(read_peers)
|
||||
|
||||
ts="$(date -u +%Y-%m-%dT%H:%M:%S.%3NZ)"
|
||||
{
|
||||
printf '{"ts":"%s","peers":[' "$ts"
|
||||
first=1
|
||||
while IFS= read -r line; do
|
||||
[ -z "$line" ] && continue
|
||||
if [ "$first" -eq 1 ]; then
|
||||
first=0
|
||||
else
|
||||
printf ','
|
||||
fi
|
||||
printf '%s' "$line"
|
||||
done <"$merge_tmp"
|
||||
printf ']}'
|
||||
} >"$AGG_TMP/cluster.json.new"
|
||||
mv "$AGG_TMP/cluster.json.new" "$CLUSTER_JSON"
|
||||
orca_log_info "aggregate" "-" "ok" "peers=$peer_count cluster_json=$CLUSTER_JSON"
|
||||
|
||||
if [ ! -d "$REMOTE_DRIFT_DIR" ] && [ ! -d "$APPLIED_DIR" ]; then
|
||||
orca_log_info "drift" "-" "skipped" "drift-events + applied dirs absent (P10b not shipped)"
|
||||
else
|
||||
drift_tmp="$AGG_TMP/drift-events"
|
||||
mkdir -p "$drift_tmp"
|
||||
drift_agg_tmp="$AGG_TMP/drift-agg.jsonl"
|
||||
: >"$drift_agg_tmp"
|
||||
while IFS= read -r peer; do
|
||||
[ -z "$peer" ] && continue
|
||||
peer_drift_dir="$drift_tmp/$peer"
|
||||
mkdir -p "$peer_drift_dir"
|
||||
if ! rsync -a --quiet $RSYNC_OPTS_RSYNC "$peer:$REMOTE_DRIFT_DIR/" "$peer_drift_dir/" 2>/dev/null; then
|
||||
orca_log_info "drift" "$peer" "skipped" "rsync failed or dir absent"
|
||||
continue
|
||||
fi
|
||||
shopt -s nullglob
|
||||
event_files=("$peer_drift_dir"/*.json)
|
||||
shopt -u nullglob
|
||||
if [ "${#event_files[@]}" -eq 0 ]; then
|
||||
orca_log_info "drift" "$peer" "ok" "no events"
|
||||
continue
|
||||
fi
|
||||
for event_file in "${event_files[@]}"; do
|
||||
event_id="$(basename "$event_file" .json)"
|
||||
path_val="$(grep -o '"path"[[:space:]]*:[[:space:]]*"[^"]*"' "$event_file" | head -1 | sed 's/.*: *"\([^"]*\)"/\1/')"
|
||||
new_sha="$(grep -o '"new_sha256"[[:space:]]*:[[:space:]]*"[^"]*"' "$event_file" | head -1 | sed 's/.*: *"\([^"]*\)"/\1/')"
|
||||
txn_val="$(grep -o '"latest_txn"[[:space:]]*:[[:space:]]*"[^"]*"' "$event_file" | head -1 | sed 's/.*: *"\([^"]*\)"/\1/')"
|
||||
status_val="$(grep -o '"status"[[:space:]]*:[[:space:]]*"[^"]*"' "$event_file" | head -1 | sed 's/.*: *"\([^"]*\)"/\1/')"
|
||||
if [ -z "$txn_val" ]; then
|
||||
orca_log_warn "drift" "$peer" "skipped" "event=$event_id no latest_txn"
|
||||
continue
|
||||
fi
|
||||
manifest="$APPLIED_DIR/$txn_val/manifest.json"
|
||||
drift_confirmed=0
|
||||
if [ -f "$manifest" ]; then
|
||||
manifest_sha="$(grep -o "\"$path_val\"[^\"]*\"sha256\"[[:space:]]*:[[:space:]]*\"[^\"]*\"" "$manifest" 2>/dev/null | grep -o '"sha256"[[:space:]]*:[[:space:]]*"[^"]*"' | head -1 | sed 's/.*: *"\([^"]*\)"/\1/')"
|
||||
if [ -n "$manifest_sha" ] && [ "$manifest_sha" != "$new_sha" ]; then
|
||||
drift_confirmed=1
|
||||
fi
|
||||
else
|
||||
orca_log_warn "drift" "$peer" "skipped" "txn=$txn_val manifest absent"
|
||||
fi
|
||||
if [ "$drift_confirmed" -eq 1 ]; then
|
||||
if [ -x "$REMEDIATE_SCRIPT" ]; then
|
||||
orca_log_info "drift" "$peer" "remediate" "txn=$txn_val path=$path_val"
|
||||
"$REMEDIATE_SCRIPT" "$peer" "$txn_val" "$path_val" \
|
||||
|| orca_log_error "drift" "$peer" "remediate-failed" "txn=$txn_val path=$path_val"
|
||||
else
|
||||
orca_log_warn "drift" "$peer" "remediate-stub" "orca-remediate.sh not installed (P10b)"
|
||||
fi
|
||||
printf '{"peer":"%s","event_id":"%s","txn":"%s","path":"%s","status":"%s","remediated":true}\n' \
|
||||
"$peer" "$event_id" "$txn_val" "$path_val" "$status_val" >>"$drift_agg_tmp"
|
||||
else
|
||||
printf '{"peer":"%s","event_id":"%s","txn":"%s","path":"%s","status":"%s","remediated":false}\n' \
|
||||
"$peer" "$event_id" "$txn_val" "$path_val" "$status_val" >>"$drift_agg_tmp"
|
||||
fi
|
||||
ssh $SSH_OPTS "$peer" "rm -f '$REMOTE_DRIFT_DIR/$event_id.json'" 2>/dev/null \
|
||||
|| orca_log_warn "drift" "$peer" "consume-failed" "event=$event_id"
|
||||
done
|
||||
done < <(read_peers)
|
||||
{
|
||||
printf '{"ts":"%s","events":[' "$ts"
|
||||
first=1
|
||||
while IFS= read -r line; do
|
||||
[ -z "$line" ] && continue
|
||||
if [ "$first" -eq 1 ]; then
|
||||
first=0
|
||||
else
|
||||
printf ','
|
||||
fi
|
||||
printf '%s' "$line"
|
||||
done <"$drift_agg_tmp"
|
||||
printf ']}'
|
||||
} >"$AGG_TMP/drift-agg.json.new"
|
||||
mv "$AGG_TMP/drift-agg.json.new" "$DRIFT_AGG_JSON"
|
||||
orca_log_info "drift" "-" "ok" "aggregated=$DRIFT_AGG_JSON"
|
||||
fi
|
||||
|
||||
orca_log_info "aggregate" "-" "done" "peers=$peer_count"
|
||||
Executable
+42
@@ -0,0 +1,42 @@
|
||||
#!/usr/bin/env bash
|
||||
# orca-watchdog.sh — lead-side watchdog meta-timer (P09, C-11).
|
||||
#
|
||||
# Runs every 30s via orca-watchdog.timer. Checks whether
|
||||
# orca-aggregate.sh has run in the last N seconds (default 30s — 3x the
|
||||
# 10s cadence). If not: fires a structured alert via
|
||||
# `logger -t orca-watchdog -p user.err "aggregator starvation detected"`.
|
||||
# The check compares the mtime of /etc/orca/state/cluster.json to the
|
||||
# current time.
|
||||
|
||||
set -euo pipefail
|
||||
|
||||
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
|
||||
# shellcheck source=lib/orca-log.sh
|
||||
. "$SCRIPT_DIR/lib/orca-log.sh"
|
||||
|
||||
ORCA_LOG_ACTOR="spiffe://orca/cli/watchdog"
|
||||
|
||||
STATE_DIR="${ORCA_STATE_DIR:-/etc/orca/state}"
|
||||
CLUSTER_JSON="$STATE_DIR/cluster.json"
|
||||
STARVATION_THRESHOLD="${ORCA_WATCHDOG_THRESHOLD:-30}"
|
||||
|
||||
alert() {
|
||||
logger -t orca-watchdog -p user.err "aggregator starvation detected: $1"
|
||||
}
|
||||
|
||||
if [ ! -f "$CLUSTER_JSON" ]; then
|
||||
alert "cluster.json absent at $CLUSTER_JSON"
|
||||
orca_log_error "watchdog" "-" "starved" "cluster.json absent"
|
||||
exit 0
|
||||
fi
|
||||
|
||||
now="$(date +%s)"
|
||||
mtime="$(stat -c %Y "$CLUSTER_JSON" 2>/dev/null || stat -f %m "$CLUSTER_JSON" 2>/dev/null || echo 0)"
|
||||
age=$((now - mtime))
|
||||
|
||||
if [ "$age" -ge "$STARVATION_THRESHOLD" ]; then
|
||||
alert "cluster.json age ${age}s >= threshold ${STARVATION_THRESHOLD}s"
|
||||
orca_log_error "watchdog" "-" "starved" "age=${age}s threshold=${STARVATION_THRESHOLD}s"
|
||||
else
|
||||
orca_log_info "watchdog" "-" "ok" "age=${age}s threshold=${STARVATION_THRESHOLD}s"
|
||||
fi
|
||||
@@ -0,0 +1,126 @@
|
||||
#!/usr/bin/env bats
|
||||
# Tests for scripts/orca-aggregate.sh + orca-watchdog.sh (P09, C-11,
|
||||
# REQ-107). Hermetic: exercises the aggregator against a fake peers.env
|
||||
# + state dir in a temp dir; the SSH path is exercised only when a
|
||||
# fake ssh shim is on PATH. The watchdog is exercised with a stale
|
||||
# cluster.json to confirm starvation detection.
|
||||
|
||||
load test_helper
|
||||
|
||||
@test "orca-aggregate.sh exists and is executable" {
|
||||
[ -f "$SCRIPTS_DIR/orca-aggregate.sh" ]
|
||||
[ -x "$SCRIPTS_DIR/orca-aggregate.sh" ]
|
||||
}
|
||||
|
||||
@test "orca-watchdog.sh exists and is executable" {
|
||||
[ -f "$SCRIPTS_DIR/orca-watchdog.sh" ]
|
||||
[ -x "$SCRIPTS_DIR/orca-watchdog.sh" ]
|
||||
}
|
||||
|
||||
@test "orca-aggregate.sh handles missing peers.env gracefully" {
|
||||
tmp="$(mktemp -d)"
|
||||
run env \
|
||||
ORCA_PEERS_ENV="$tmp/nonexistent.env" \
|
||||
ORCA_STATE_DIR="$tmp/state" \
|
||||
ORCA_REMOTE_STATE_DIR="$tmp/remote-state" \
|
||||
ORCA_REMOTE_DRIFT_DIR="$tmp/remote-drift" \
|
||||
ORCA_APPLIED_DIR="$tmp/applied" \
|
||||
bash "$SCRIPTS_DIR/orca-aggregate.sh"
|
||||
assert_status 0 "$status"
|
||||
# cluster.json written atomically with an empty peers array.
|
||||
[ -f "$tmp/state/cluster.json" ]
|
||||
assert_contains "$(cat "$tmp/state/cluster.json")" '"peers":[]'
|
||||
}
|
||||
|
||||
@test "orca-aggregate.sh skips drift section when drift-events absent" {
|
||||
tmp="$(mktemp -d)"
|
||||
# No applied dir and no remote drift dir on lead → drift section
|
||||
# must skip (no drift-events-aggregated.json written).
|
||||
run env \
|
||||
ORCA_PEERS_ENV="$tmp/nonexistent.env" \
|
||||
ORCA_STATE_DIR="$tmp/state" \
|
||||
ORCA_REMOTE_STATE_DIR="$tmp/remote-state" \
|
||||
ORCA_REMOTE_DRIFT_DIR="$tmp/absent-drift" \
|
||||
ORCA_APPLIED_DIR="$tmp/absent-applied" \
|
||||
bash "$SCRIPTS_DIR/orca-aggregate.sh"
|
||||
assert_status 0 "$status"
|
||||
[ -f "$tmp/state/cluster.json" ]
|
||||
# Drift aggregation skipped → aggregated file must NOT exist.
|
||||
[ ! -f "$tmp/state/drift-events-aggregated.json" ]
|
||||
}
|
||||
|
||||
@test "orca-watchdog.sh detects starvation with old cluster.json" {
|
||||
tmp="$(mktemp -d)"
|
||||
mkdir -p "$tmp/state"
|
||||
cluster="$tmp/state/cluster.json"
|
||||
echo '{}' >"$cluster"
|
||||
# Set mtime to 120s ago so age (120) >= threshold (default 30).
|
||||
old_ts="$(date -d '120 seconds ago' +%Y%m%d%H%M.%S)"
|
||||
touch -t "$old_ts" "$cluster"
|
||||
# logger is available in CI; capture the alert by overriding PATH
|
||||
# with a fake logger that records to a file.
|
||||
fake_bin="$tmp/bin"
|
||||
mkdir -p "$fake_bin"
|
||||
cat >"$fake_bin/logger" <<'LOGGER_EOF'
|
||||
#!/usr/bin/env bash
|
||||
echo "logger: $*" >>"$LOGGER_OUT"
|
||||
LOGGER_EOF
|
||||
chmod +x "$fake_bin/logger"
|
||||
env \
|
||||
ORCA_STATE_DIR="$tmp/state" \
|
||||
ORCA_WATCHDOG_THRESHOLD=30 \
|
||||
LOGGER_OUT="$tmp/logger.out" \
|
||||
PATH="$fake_bin:$PATH" \
|
||||
bash "$SCRIPTS_DIR/orca-watchdog.sh"
|
||||
[ -f "$tmp/logger.out" ]
|
||||
out="$(cat "$tmp/logger.out")"
|
||||
assert_contains "$out" "aggregator starvation detected"
|
||||
}
|
||||
|
||||
@test "orca-watchdog.sh passes when cluster.json is fresh" {
|
||||
tmp="$(mktemp -d)"
|
||||
mkdir -p "$tmp/state"
|
||||
echo '{}' >"$tmp/state/cluster.json"
|
||||
# Fresh mtime (now); age 0 < threshold.
|
||||
fake_bin="$tmp/bin"
|
||||
mkdir -p "$fake_bin"
|
||||
cat >"$fake_bin/logger" <<'LOGGER_EOF'
|
||||
#!/usr/bin/env bash
|
||||
echo "logger: $*" >>"$LOGGER_OUT"
|
||||
LOGGER_EOF
|
||||
chmod +x "$fake_bin/logger"
|
||||
env \
|
||||
ORCA_STATE_DIR="$tmp/state" \
|
||||
ORCA_WATCHDOG_THRESHOLD=30 \
|
||||
LOGGER_OUT="$tmp/logger.out" \
|
||||
PATH="$fake_bin:$PATH" \
|
||||
bash "$SCRIPTS_DIR/orca-watchdog.sh"
|
||||
if [ -f "$tmp/logger.out" ]; then
|
||||
out="$(cat "$tmp/logger.out")"
|
||||
! assert_contains "$out" "aggregator starvation detected" || {
|
||||
echo "unexpected starvation alert for fresh cluster.json" >&2
|
||||
return 1
|
||||
}
|
||||
fi
|
||||
}
|
||||
|
||||
@test "orca-watchdog.sh alerts when cluster.json absent" {
|
||||
tmp="$(mktemp -d)"
|
||||
fake_bin="$tmp/bin"
|
||||
mkdir -p "$fake_bin"
|
||||
cat >"$fake_bin/logger" <<'LOGGER_EOF'
|
||||
#!/usr/bin/env bash
|
||||
echo "logger: $*" >>"$LOGGER_OUT"
|
||||
LOGGER_EOF
|
||||
chmod +x "$fake_bin/logger"
|
||||
env \
|
||||
ORCA_STATE_DIR="$tmp/state" \
|
||||
ORCA_WATCHDOG_THRESHOLD=30 \
|
||||
LOGGER_OUT="$tmp/logger.out" \
|
||||
PATH="$fake_bin:$PATH" \
|
||||
bash "$SCRIPTS_DIR/orca-watchdog.sh"
|
||||
[ -f "$tmp/logger.out" ]
|
||||
out="$(cat "$tmp/logger.out")"
|
||||
assert_contains "$out" "aggregator starvation detected"
|
||||
assert_contains "$out" "cluster.json absent"
|
||||
}
|
||||
Reference in New Issue
Block a user