Files
fancy-dns/AGENTS.md
T
developer b6d70a50fe
build / build-push (push) Successful in 22s
Finish the server shutdown tests; fix the two bugs they found
internal/server was the last package whose shutdown was tested by assumption:
four transports, three shutdown shapes, and every one of them can strand a
goroutine that a client is holding open. Both bugs below were live in the tree
as handed over and neither had a test.

An idle client could hold the whole shutdown open. The stream accept loop waits
for its connections before Stop returns, and a connection parked in ReadMsg with
no deadline in front of it only ends when the client goes away — ctx.Err() is
checked between reads, so cancelling the context cannot reach a read that is
already blocked. One connection that connected and said nothing was enough to
make Stop sit there for the caller's entire grace period (15s in main) and then
return context.DeadlineExceeded, which the app reports as "shutdown completed
with errors"; and the cache snapshot the app takes right after Stop never
happened, because the orchestrator had given up first. It is a shutdown DoS that
costs the attacker one connection. The fix is a read deadline in the past on
every accepted connection when the context is cancelled, and deliberately not a
close: a query already inside the pipeline is past its read, so its answer still
goes out, which is the invariant TestInFlightQueryIsAnsweredAcrossStop pins. The
accept loop keeps a small registry of live connections for this — the accept
path, not the query path, and no query ever takes that mutex.

A UDP query in flight at Stop was lost and nothing waited for it. The stream
paths register their connections with s.wait; servePacket did not register at
all, so a UDP query already inside the pipeline was gone twice over: nothing
waited for it, and the answer had nowhere to go once it was ready, because a
datagram is written to the very socket the query arrived on. The visible damage
was not the lost answer (the client retries) but the app, which snapshots the
cache and closes the store the moment Stop returns — a query still in the
pipeline could be writing to a store on its way out. Packets are now counted in
s.wait, and the UDP listener's close wakes the read loop with an expired
deadline and then waits, bounded by udpDrainGrace (250ms), for the loop and the
packets it spawned before closing the socket. The bound is deliberate: a
shutdown may not spend its grace period on one slow upstream, and past the bound
the answer is dropped, which is what the client would see anyway.

Stop returning did not mean the process was done, so nothing could tell a
leaking shutdown from a clean one.
TestStopLeavesNoGoroutineRunningServerCode counts goroutines whose stack still
has a frame in the package — not runtime.NumGoroutine(), which in a test binary
is net/http, crypto/tls and the race detector talking over each other — and it
found the idle-connection bug above. TestStopIsNotBlockedByAnIdleStreamConnection
sets a 60s DoT idle timeout on purpose so a config value cannot hide the
regression behind a lucky tick, and TestInFlightQueryIsAnsweredAcrossStop now
covers all four transports. harness.idleStream exists because a test that only
dials a stream transport races the accept loop: Stop can close the listener
while the connection is still in the backlog, the server never accepts it, and a
shutdown assertion that needs a live server-side goroutine passes for the wrong
reason.

AGENTS.md gains the two durable invariants (no step in the shutdown may wait for
a client; every query goroutine is registered with s.wait, UDP included) and the
two test-hygiene traps. STATUS.md drops the UDP query lost at Stop from its
pinned list, since it is now a test rather than an accident, and the plan starts
at the pinned decisions that are actually left.
2026-09-27 19:28:40 +00:00

13 KiB

AGENTS.md — working on fancy-dns

Project-specific knowledge for anyone (human or agent) changing this repo. General state and the continuation plan live in STATUS.md; read that first for what is done and what is next. This file is the part that should not change every week: the decisions, the invariants, and the rules that have been learned the hard way.

What this is

A recursive DNS server in Go with an embedded React console. Config is YAML plus RFC 1035 master files, hot-reloaded over SIGHUP. Blocklists in four formats, a sharded LRU cache with a bbolt snapshot for warm restart, per-client ACLs, query logging, Prometheus metrics, and optional Valkey/Redis for cross-instance sync.

Layout

cmd/fancy-dns/main.go          entrypoint: flags, signals, SIGHUP reload
internal/
  config/       Config struct, Duration, Size, Validate, atomic Store + hooks,
                ${VAR} environment expansion in Load
  zonefile/     RFC 1035 master-file parser ($TTL/$ORIGIN, parens, serial math)
  blocklist/    Set (exact/sub/root maps), 4 parsers, Manager (fetch+compile+swap)
  cache/        Key, Entry, sharded LRU, TTL policy, serve-stale, bbolt Store
  resolver/     Upstream pool: health, circuit breakers, parallel/failover/random
  acl/          compiled groups, CIDR/PTR/tag matching, per-group domain policy
  localzone/    authoritative data from master files, wildcards, delegations
  engine/       THE QUERY PIPELINE, reverse-PTR cache
  server/       UDP, TCP, DoT, DoH listeners; all funnel into one handler
  qlog/         bbolt query-log ring + live in-memory ring + SSE fanout
  metrics/      Prometheus collectors + rolling time-series rings
  cluster/      Valkey pub/sub, leader lease, divergence detection
  api/          admin REST + SSE, auth, CORS, rate limiting
  app/          wiring, lifecycle, API hook implementations, memory envelope
  web/          go:embed of the built console
web/            Vite + React + TS + Tailwind console
deploy/         example config, example zones, k8s manifests

Query pipeline order (engine.Handle)

validate → ACL deny → per-client concurrency → global concurrency → rate limit
  → group allow/deny domains → local zone → blocklist → cache → upstream
  → cache put → finalize (EDNS0, UDP truncation) → metrics + qlog

Cache is a single shard-local lookup; only a miss touches the network.

Decisions already made — do not relitigate

Decision Choice Why
Scope Full production clustering, metrics, ACLs, query logs, hot reload, Docker, k8s, CI
Config store YAML + DNS master files reviewable/diffable in git
Cluster sync Valkey (Redis protocol) simple, mature Go client, easy local dev
Secure DNS DoT + DoH, no DNSSEC validation deliberately out of scope; AD/CD/extended-RCODE are passed through only
Hot cache Sharded LRU in memory, bbolt for snapshots no I/O on the hot path; warm restart
k8s networking No hostNetwork, LoadBalancer Service the server never sets SO_REUSEPORT, so two host-network pods cannot share :53
Memory Soft GOMEMLIMIT + cache byte budget a cgroup kill is an outage; a cache miss is not
Scale-out StatefulSet, no HPA an HPA cannot drive a StatefulSet; scale by adding replicas
Ingress A separate ClusterIP Service a headless Service has no virtual IP for an Ingress to route to
Profiling pprof refused unless admin is loopback the handlers sit outside the API's auth middleware

Invariants

Break these and you break the design, not just the test.

  • Everything the query path reads is behind an atomic.Pointer holder — ACL, blocklist, localzone, upstream pool. Readers load the pointer once and work against a value that cannot change under them. A lock on the hot path is a regression. Prefer building a new immutable snapshot and swapping the pointer.
  • Cache entries are packed wire bytes, not *dns.Msg. A hit costs one Unpack and yields a private copy, so serving never mutates shared state and TTL ageing is free. Same trade dnsmasq makes.
  • The cache is bounded by bytes as well as by entry count. cache.size alone bounds nothing: one cached ANY or TXT answer can be tens of kilobytes.
  • Serve-stale answers are served with TTL 1. An expired upstream answer is never cached. A truncated answer is never cached — it would be served truncated forever.
  • The UDP read deadline is reset every iteration. A deadline set once expires for good and turns the loop into a spin that never reads another packet.
  • No step in the shutdown may wait for a client. Stop is: cancel, close every listener (not interruptible — releasing the ports is the work the next process start needs), then wait for the goroutines (interruptible by the caller's context). The wait is the step that breaks when something depends on a client: a connection parked in a read with no deadline in front of it never notices the cancellation, and one idle connection is enough to spend the whole grace period and then report a context error. Reach those connections with a read deadline in the past, not a close — a query already inside the pipeline is past its read, and its answer still has to go out.
  • Every query goroutine is registered with s.wait, UDP included, so Stop returning means nothing is still inside the engine: the app snapshots the cache and closes the store immediately afterwards, and a query still in flight would be writing to a store on its way out. A UDP answer has nowhere else to go, so that transport also gets a short bounded drain before its socket closes — see udpDrainGrace in internal/server/server.go.
  • Only the Valkey lease holder fetches blocklists, so N instances do not hit a mirror N times. Followers serve their compiled copy.
  • Blocklist changes flush the cache rather than guessing which subtrees are affected (invalidateBlocklistCache() in internal/app/hooks.go).
  • API keys are compared by SHA-256 hash, constant-time, and X-Forwarded-For is honoured only from trusted_proxies. Honouring it from anyone lets a client spoof its identity past the ACL and the rate limiter.
  • The API response field names are shared with the frontend. If you rename one, grep web/src too.
  • PatternKind.Pattern is the display form; WireName is the wildcard form. The internal *. / **. prefixes must not leak into the UI, but the API needs them to re-add a rule unchanged.
  • The Go heap does not account for goroutine stacks, runtime metadata, or the bbolt mmap. memory.reserve exists to cover them; the cache snapshot and query log are mmapped and live outside the heap entirely.
  • Unknown config keys are silently ignored (a settled decision, pinned by TestLoadIgnoresUnknownKeys in internal/config: yaml KnownFields(true) was considered and rejected as a behaviour change for anyone carrying extra keys). If a config key seems to have no effect, check the spelling against the struct tags before anything else. The leniency covers unknown keys only — a known key with a bad type is still a startup failure.
  • dns.EDNS0_NSID.Nsid is hex on the wire. A human-readable value there builds a message that will not pack, so the client gets SERVFAIL. See the test-hygiene traps below.

Working rules

Never leave the tree red. You have a hard budget and no way to detect you are close to it. Whatever is on disk at the end of any round must compile and pass. Get a few tests running early and keep them running. If you are running long, stop adding tests and make the existing ones pass — a smaller green suite is worth far more than a larger red one.

Scope your edits. Check git status before you start; other work may be in flight. Stage only the packages you touched. .codesearch/ is tracked but gitignored — never stage it. If you edit a package someone else is also editing, say so in your report rather than silently overwriting.

Verify before you claim. make check-go is the fast loop (gofmt, vet, compile, ~1s). make check adds the full test suite. Push only a green tree.

Commit and push when the task says so. Read git log --oneline -5 and git show --stat HEAD first and match the style: an imperative subject line, then one paragraph per bug covering root cause and fix. A commit that says what was changed without saying why is not finished work.

When a test fails, decide which of three things is wrong before editing anything: the test, the implementation, or the specification. All three happen here. Say which one you concluded and why. Do not relax an assertion to make a suite pass — if an existing assertion pins buggy behaviour, change it and say so explicitly in the commit message, because that decision belongs in the history.

Report honestly. If you did not finish, say what is unfinished and what state the tree is in. A partial result that is accurately described is usable; a partial result described as done is not.

Test-hygiene traps

  • dns.NewRR needs a TTL column: dns.NewRR("host.example.com") does not parse. A failure there is a bad fixture, not a product bug.
  • A test that closes a stop channel after wg.Wait() deadlocks the whole package and looks like "the suite hangs". Close first, then wait.
  • Assert on invariants, not on what a concurrent goroutine happened to observe. "A reader must have seen address X" is a flake; "every list a reader sees is one of the published ones" is a real property and cannot flake.
  • A test that only dials a stream transport races the accept loop. A shutdown can close the listener while the connection is still in the backlog, so the server never accepts it and the test passes for the wrong reason. Exchange one query first, then leave the connection idle (harness.idleStream).
  • Count leaked goroutines by stack, not by runtime.NumGoroutine(). The test binary shares its process with net/http, crypto/tls and the race detector, so a baseline count is noise. "Is any goroutine still running code in this package?" is a property of the product — and it needs the caller's own goroutine filtered out by id, or every run finds exactly one.
  • Spin up dns.Server fakes on 127.0.0.1:0: net.Listen/net.ListenPacket to find the port, ActivateAndServe in a goroutine, stop in t.Cleanup.
  • A dns.Msg carrying a TXT string over 255 bytes does not pack in miekg/dns v1.1.73 — the error is dns: string exceeded 255 bytes in txt, and 255 is the exact boundary. On a fake upstream's write path that failure is silent and presents as a client timeout or a SERVFAIL several layers away. Build a multi-kilobyte answer out of many strings under 255 bytes, and assert the fake's write error before blaming the client.
  • dns.EDNS0_NSID.Nsid is a hex string, not the bytes on the wire. The library hex-decodes it in pack and hex-encodes it in unpack, so a plain value such as "fancy-dns" fails with dns: overflow packing opt and a long NSID packs fine (unlike TXT). This bit the engine: an unencoded literal made every query from an NSID-using client SERVFAIL. Always hex-encode on the way in; TestPinnedEDNS0NSIDIsHex in internal/engine/pins_test.go pins the contract.
  • internal/app/integration_test.go already builds a fully wired app. Reuse its harness rather than rebuilding one.
  • tsc -b replays cached diagnostics from node_modules/.tmp/*.tsbuildinfo. After a large frontend edit: rm node_modules/.tmp/*.tsbuildinfo.
  • vite build empties outDir and deletes internal/web/dist/.gitkeep. Restore it or the directory disappears from git.
  • The Go binary embeds internal/web/dist, so npm ci && npm run build must run before a fresh clone will build a binary with the console in it.
  • make build depends on web and shells out to npm. make go-build skips the console and is what you want in an inner loop.

Security-sensitive areas

These have been wrong before. Changes here need a test that fails without the fix.

  • Auth (internal/api/auth.go): API key comparison must stay constant-time; the key is hashed before comparison so the plaintext is not in a long-lived string.
  • Client identity: X-Forwarded-For is only trusted from trusted_proxies, and httpRemoteAddr (used by the DoH handler) deliberately does not consult it at all. Behind an Ingress every DoH client therefore lands in the proxy's ACL group — fail-closed if that group is restrictive, fail-open if it is permissive. Pin this behaviour with a test before changing it.
  • Config expansion: ${VAR} is expanded in Load and an unset variable with no default is fatal, so a missing Secret is a crash loop rather than an unauthenticated console. The expansion works inside quoted scalars — that is the form operators write — and skips YAML comments. Do not "simplify" it into a plain os.ExpandEnv; that silently breaks the API key path.
  • pprof is loopback-only by construction, so no config flag can turn it into a network-reachable unauthenticated endpoint. Keep it that way.
  • Amplification: the DoH/UDP handlers cap the client's advertised EDNS0 buffer. Do not remove the cap.