KB-002: Hyperdrive Binding Connected but database Service Reports down

Status: ✅ Active Affected version: v0.75.0 Resolved in: PR fixing PrismaClientConfigSchema to accept postgres:// + enhanced /api/health probe Date: 2026-03-25

Session Log — 2026-03-25

This section captures the live troubleshooting session that led to the discovery of this KB article and the subsequent hardening work in v0.76.0+.

Symptoms Observed

• UI showed "Degraded performance — v0.75.0" and "Data may be stale" banners on every page load
• /api/health returned database.status: "down" with latency_ms: 0
• Cloudflare Hyperdrive admin page showed zero traffic despite Neon dashboard showing migration activity
• Health response showed hyperdrive_host: "<resource-id>.hyperdrive.local" — which is the correct Hyperdrive local proxy address (not a bug)

Diagnosis Steps Taken

# Step 1: Inspect the full health response
curl -s https://app.bloqr.dev/api/health | jq .services.database
# Result:
# {
#   "status": "down",
#   "latency_ms": 0,
#   "hyperdrive_host": "<resource-id>.hyperdrive.local"
# }

# Step 2: Check the Hyperdrive binding configuration
wrangler hyperdrive get <resource-id>
# Result showed "scheme": "postgres" — Hyperdrive uses postgres://, not postgresql://

# Step 3: Check deployed version
curl -s https://app.bloqr.dev/api/health | jq .version
# "0.76.0" — schema fix was deployed

# Step 4: Tail live worker logs
wrangler tail --format=pretty
# Revealed: ZodError thrown before any network call when parsing connectionString

Key Diagnostic Insight: .hyperdrive.local Is Correct

The hyperdrive_host: "<resource-id>.hyperdrive.local" host is the Cloudflare-managed local proxy address inside a deployed Worker. This is expected and correct — it means the HYPERDRIVE binding is wired up properly. The failure was happening before any query reached this proxy.

Root Cause Found

PrismaClientConfigSchema in worker/lib/prisma-config.ts was only accepting postgresql:// as a valid scheme. Cloudflare Hyperdrive returns postgres:// from env.HYPERDRIVE.connectionString. This caused a ZodError to be thrown synchronously before any TCP connection was attempted — hence latency_ms: 0.

The fix (accepting both schemes via .regex(/^postgre(?:s|sql):\/\//)) was already applied in the first PR.

Remaining Issue at v0.76.0

Even after the schema fix was deployed at v0.76.0, the database was still down. The latency_ms: 0 pattern continued, meaning the error was still at instantiation time, not at query time. At this point there were two hypotheses:

1. The schema fix wasn't actually deployed — wrangler deployments list should confirm the version
2. Query-level failure — the PrismaClient was being created but the query itself was failing silently (catch block swallowed the error)

The original databaseProbe catch block did not capture error_code or error_message, making it impossible to distinguish these cases from the health response alone.

Resolution Path

The hardening work in this PR added:

1. error_code and error_message in the databaseProbe catch block — redacted of any connection string fragments
2. 5-second timeout via Promise.race — a hung Hyperdrive connection no longer blocks the health response indefinitely
3. $disconnect() in a finally block — prevents connection pool leaks after each probe
4. New GET /api/health/db-smoke endpoint — runs current_database(), pg_version, now(), and table_count as a richer smoke test. This is the canonical way to verify DB connectivity after every production deploy.

Using the New Smoke Test Endpoint

After deploying, run:

curl -s https://app.bloqr.dev/api/health/db-smoke | jq .

Expected healthy output:

{
  "ok": true,
  "db_name": "bloqr-backend",
  "pg_version": "PostgreSQL 16.x ...",
  "server_time": "2026-03-25T21:59:15.917Z",
  "table_count": 17,
  "latency_ms": 42,
  "hyperdrive_host": "ep-xxx-pooler.eastus2.azure.neon.tech"
}

If it returns ok: false, the error field will now contain a redacted error message that pinpoints the failure layer.

Symptom

The live site at https://app.bloqr.dev/ displays two error banners:

• "Degraded performance — v0.75.0"
• "Data may be stale"

Hitting the health endpoint returns:

{
  "status": "down",
  "version": "0.75.0",
  "timestamp": "2026-03-25T21:59:15.917Z",
  "services": {
    "gateway":  { "status": "healthy" },
    "database": { "status": "down", "latency_ms": 0 },
    "compiler": { "status": "healthy" },
    "auth":     { "status": "healthy", "provider": "better-auth" },
    "cache":    { "status": "healthy", "latency_ms": 132 }
  }
}

Key tell: latency_ms: 0 on the database service. A real network failure or timeout always returns a non-zero latency. An instant 0 ms failure means the probe threw before any connection attempt — i.e., at the validation layer.

The Cloudflare Hyperdrive dashboard shows zero queries/connections despite the Neon dashboard showing migration activity.

Diagnostic Commands

# 1. Inspect the full health response
curl -s https://<your-worker>.workers.dev/api/health | jq .

# 2. Check the Hyperdrive binding configuration
npx wrangler hyperdrive get <hyperdrive-id>

# 3. Tail the live worker log to catch Zod validation errors
wrangler tail

Look for lines like ZodError: Invalid url or Expected string, received undefined in the tail output. A ZodError thrown during PrismaClientConfigSchema.parse() is definitive proof that the failure is at the config-validation layer, not the network.

Root Cause Decision Tree

❶ Is latency_ms exactly 0?

If YES — the database probe threw before opening any connection. This points to a config-validation failure, not a network failure. Proceed to ❷.

If NO (latency is non-zero) — the connection was attempted but timed out or was refused. Check Neon project status, Hyperdrive binding ID, and network egress. This article does not cover that case.

❷ What scheme does the Hyperdrive connection string use?

npx wrangler hyperdrive get <hyperdrive-id>

The output will show a "scheme" field. Cloudflare Hyperdrive always returns postgres:// (not postgresql://) from the env.HYPERDRIVE.connectionString binding property.

{
  "id": "...",
  "name": "adblock-hyperdrive",
  "origin": {
    "scheme": "postgres",
    "host": "...",
    "port": 5432,
    "database": "neondb"
  }
}

If the scheme is postgres — and PrismaClientConfigSchema only accepts postgresql://, the schema parse throws a ZodError instantly. This is the root cause. Proceed to Resolution.

❸ Does PrismaClientConfigSchema accept postgres://?

Open worker/lib/prisma-config.ts and check the connectionString validator:

// Before the fix — rejects Hyperdrive's actual scheme
connectionString: z.string().url().startsWith('postgresql://'),

// After the fix — accepts both schemes
connectionString: z.string().url().regex(/^postgre(?:s|sql):\/\//),

If the schema only allows postgresql://, every request that tries to build a PrismaClient from the Hyperdrive binding will fail at parse time with zero network activity.

Resolution

Step 1 — Update PrismaClientConfigSchema to accept both URL schemes

In worker/lib/prisma-config.ts, change the connectionString validation to accept both postgres:// and postgresql://:

// worker/lib/prisma-config.ts

// Before
connectionString: z.string().url().startsWith('postgresql://'),

// After
connectionString: z.string().url().regex(/^postgre(?:s|sql):\/\//),

This accepts postgres://... (Hyperdrive short alias) and postgresql://... (standard long form) while rejecting anything else.

Step 2 — Deploy

wrangler deploy

After deploying, hit the health endpoint again:

curl -s https://<your-worker>.workers.dev/api/health | jq .

You should see database.status: "healthy" with a non-zero latency_ms.

Step 3 — Verify with the enhanced health probe

The fix also introduced an enhanced health check in worker/handlers/health.ts that runs SELECT current_database() instead of SELECT 1. This surfaces db_name and hyperdrive_host in the health response:

{
  "services": {
    "database": {
      "status": "healthy",
      "latency_ms": 42,
      "db_name": "neondb",
      "hyperdrive_host": "...-pooler.us-east-2.aws.neon.tech"
    }
  }
}

Confirm that db_name matches the expected Neon database name. If it returns a different database name, the Hyperdrive binding is pointed at the wrong Neon project or branch.

Prevention

• The new db_name field in the health response acts as a continuous assertion that the correct database is connected. Monitor this value in your observability dashboards.
• When configuring a new Hyperdrive binding, always run wrangler hyperdrive get <id> to confirm the scheme field. If it is "postgres", ensure all Zod schemas that validate the connection string accept postgres://.
• Add an integration test that builds PrismaClientConfigSchema.parse() with a postgres:// URL to catch future regressions.

Worker Code Reference

ZTA Security Note

env.HYPERDRIVE.connectionString is a runtime binding secret — it is never logged, committed, or exposed in the health response. Only the host portion (hyperdrive_host) is surfaced for diagnostic purposes. The db_name field is also safe to expose: it is a non-secret label, not a credential.

Resolution Summary

Related KB Articles

• KB-001 — "Getting API is not available" on the main page
• KB-003 — Database Down After Deploy — Live Debugging Session (2026-03-25)
• KB-004 — WebAssembly.Module(): Wasm code generation disallowed by embedder — Prisma runtime = "deno" generates code blocked by Cloudflare Workers; fix is runtime = "cloudflare" + regenerate
• KB-005 — Better Auth Cloudflare integration issues: Worker CPU timeout on session endpoints and silent rate limiting bypass
• (planned) KB-006 — Cloudflare Queue consumer not processing messages
• (planned) KB-007 — Angular SPA serves stale build after worker deploy

Feedback & Contribution

If you discovered a new failure mode while using this article, please open an issue tagged troubleshooting and documentation in bloqr-systems/bloqr-backend with the details so it can be captured in a follow-up KB entry.