Fix two bugs found by actually running the Phase 0 pipeline end-to-end

Both surfaced only by running docker compose up for real, not from review:

- ClickHouse's official image silently disables network access for the
  default user unless CLICKHOUSE_USER or CLICKHOUSE_PASSWORD is set to a
  genuinely non-empty value (an explicit empty password still triggers
  it). Set a dev-only password across clickhouse, clickhouse-migrate,
  ingest, and api in both docker-compose.yml files.

- rpk cluster health and rpk topic ... don't accept --brokers; health
  checks need -X admin.hosts=... (port 9644), topic commands need
  -X brokers=... (port 9092). The old script's retry loop silently
  swallowed the resulting "unknown flag" error and retried forever,
  which blocked ingest from ever starting.

Verified: agent -> ingest -> Redpanda -> ClickHouse -> api round-trip
confirmed with a real log line on a real host.

docker-compose.yml

@@ -43,6 +43,7 @@ services:
         condition: service_healthy
     environment:
       REDPANDA_BROKERS: "redpanda:9092"
+      REDPANDA_ADMIN_HOSTS: "redpanda:9644"
 
   clickhouse:
     image: clickhouse/clickhouse-server:24.8
@@ -50,6 +51,17 @@ services:
     ports:
       - "8123:8123" # HTTP interface, used by the migrate step
       - "9000:9000" # native protocol, used by ingest and api
+    environment:
+      # The official image disables *network* access entirely for the
+      # default user (even from sibling containers) unless
+      # CLICKHOUSE_USER or CLICKHOUSE_PASSWORD is set to a genuinely
+      # non-empty value — confirmed by testing, not just reading docs: an
+      # explicitly-empty CLICKHOUSE_PASSWORD="" still triggers the
+      # lockdown, silently returning 403 to every other container. This
+      # password isn't a real secret (mTLS between agent and ingest is
+      # the actual security boundary here) — it exists purely to satisfy
+      # this image's login gate for local/homelab use.
+      CLICKHOUSE_PASSWORD: "sentry-dev-only"
     volumes:
       - clickhouse-data:/var/lib/clickhouse
     ulimits:
@@ -73,6 +85,7 @@ services:
         condition: service_healthy
     environment:
       CLICKHOUSE_HTTP: "http://clickhouse:8123"
+      CLICKHOUSE_PASSWORD: "sentry-dev-only"
 
   ingest:
     build:
@@ -89,6 +102,7 @@ services:
     environment:
       REDPANDA_BROKERS: "redpanda:9092"
       CLICKHOUSE_ADDR: "clickhouse:9000"
+      CLICKHOUSE_PASSWORD: "sentry-dev-only"
       # TLS_*_FILE env vars are left at their defaults
       # (/etc/sentry-ingest/{server,server-key,ca}.pem) — matches where
       # the volume below mounts the generated dev certs.
@@ -107,6 +121,7 @@ services:
       - "8080:8080"
     environment:
       CLICKHOUSE_ADDR: "clickhouse:9000"
+      CLICKHOUSE_PASSWORD: "sentry-dev-only"
 
   web:
     build:

storage/README.md

@@ -59,18 +59,28 @@ split multi-statement files — keep each migration to a single statement.
 ## Running
 
 ```sh
-docker compose up -d          # starts a standalone ClickHouse for local work
+docker compose up -d                          # starts a standalone ClickHouse for local work
-./migrate.sh                  # applies migrations/*.sql
+CLICKHOUSE_PASSWORD=sentry-dev-only ./migrate.sh   # applies migrations/*.sql
 ```
 
-Environment variables `migrate.sh` reads (all optional, matching
+`CLICKHOUSE_PASSWORD` here has to match whatever `docker-compose.yml` set
-`/ingest`'s ClickHouse defaults so the two stay in sync out of the box):
+on the `clickhouse` service — found this the hard way running the Phase 0
+runbook for real: the official ClickHouse image silently disables *all*
+network access (including the published port, not just container-to-
+container traffic) for the `default` user unless `CLICKHOUSE_USER` or
+`CLICKHOUSE_PASSWORD` is a genuinely non-empty value. An empty
+`CLICKHOUSE_PASSWORD=""` still triggers the lockdown — it has to actually
+have a value. Not a real secret, just what this image demands.
+
+Environment variables `migrate.sh` reads (all optional except
+`CLICKHOUSE_PASSWORD` as of the above, matching `/ingest`'s ClickHouse
+defaults so the two stay in sync out of the box):
 
 | Var | Default |
 |---|---|
 | `CLICKHOUSE_HTTP` | `http://localhost:8123` |
 | `CLICKHOUSE_USER` | `default` |
-| `CLICKHOUSE_PASSWORD` | (empty) |
+| `CLICKHOUSE_PASSWORD` | (empty — override, see above) |
 | `CLICKHOUSE_DATABASE` | `sentry` |
 
 There's also a `Dockerfile` (bash + curl baked in, `migrations/` copied in

storage/docker-compose.yml

@@ -9,6 +9,15 @@ services:
     ports:
       - "8123:8123" # HTTP interface, used by migrate.sh
       - "9000:9000" # native protocol, used by ingest
+    environment:
+      # The official image disables *network* access entirely for the
+      # default user (this includes the published port, not just
+      # container-to-container traffic) unless CLICKHOUSE_USER or
+      # CLICKHOUSE_PASSWORD is set to a genuinely non-empty value — an
+      # explicitly-empty CLICKHOUSE_PASSWORD="" still triggers it. Not a
+      # real secret; see the root docker-compose.yml for the full
+      # explanation.
+      CLICKHOUSE_PASSWORD: "sentry-dev-only"
     volumes:
       - clickhouse-data:/var/lib/clickhouse
     ulimits:

transport/README.md

@@ -15,9 +15,17 @@ both are invoked.
 
 ```sh
 docker compose up -d
-REDPANDA_BROKERS=localhost:9092 ./provision-topics.sh
+./provision-topics.sh   # defaults (localhost:9092 / localhost:9644) match this compose file
 ```
 
+Two separate addresses matter here, confirmed by actually running this
+against a live Redpanda container: `rpk cluster health` talks to the
+**Admin API** (`REDPANDA_ADMIN_HOSTS`, port 9644), while `rpk topic ...`
+talks to the **Kafka API** (`REDPANDA_BROKERS`, port 9092) — and neither
+accepts a `--brokers` flag directly, both need `-X admin.hosts=...` /
+`-X brokers=...`. Get this wrong and it doesn't error loudly: it just
+retries the health check forever without ever reporting why.
+
 ## In the full stack
 
 The root-level `docker-compose.yml` builds this directory's `Dockerfile`

transport/provision-topics.sh

@@ -6,18 +6,28 @@
 # sibling container on the root compose's network, or in CI.
 set -euo pipefail
 
+# NOTE: confirmed by actually running this against a live Redpanda
+# container -- neither `rpk cluster health` nor `rpk topic ...` accept a
+# `--brokers` flag in this rpk version. Health checks hit the Admin API
+# (-X admin.hosts=..., port 9644); topic commands hit the Kafka API
+# (-X brokers=..., port 9092). Getting this wrong doesn't error loudly:
+# `rpk cluster health --brokers ...` fails with "unknown flag" but that
+# failure was swallowed by this script's own `> /dev/null 2>&1` retry
+# loop, which just silently retried the malformed command forever instead
+# of ever becoming healthy.
 BROKERS="${REDPANDA_BROKERS:-localhost:9092}"
+ADMIN_HOSTS="${REDPANDA_ADMIN_HOSTS:-localhost:9644}"
 TOPIC="${REDPANDA_TOPIC:-sentry.logs.raw}"
 PARTITIONS="${REDPANDA_TOPIC_PARTITIONS:-6}"
 
-echo "Waiting for Redpanda at ${BROKERS}..."
+echo "Waiting for Redpanda admin API at ${ADMIN_HOSTS}..."
-until rpk cluster health --brokers "${BROKERS}" --exit-when-healthy > /dev/null 2>&1; do
+until rpk cluster health -X "admin.hosts=${ADMIN_HOSTS}" --exit-when-healthy > /dev/null 2>&1; do
     sleep 1
 done
 
-if rpk topic list --brokers "${BROKERS}" | awk 'NR>1{print $1}' | grep -qx "${TOPIC}"; then
+if rpk topic list -X "brokers=${BROKERS}" | awk 'NR>1{print $1}' | grep -qx "${TOPIC}"; then
     echo "Topic '${TOPIC}' already exists, skipping."
 else
     echo "Creating topic '${TOPIC}' (${PARTITIONS} partitions)..."
-    rpk topic create "${TOPIC}" --brokers "${BROKERS}" --partitions "${PARTITIONS}" --replicas 1
+    rpk topic create "${TOPIC}" -X "brokers=${BROKERS}" --partitions "${PARTITIONS}" --replicas 1
 fi