Files
bucketeer/ARCHITECTURE.md
T
StasonJatham 8fce55f8c9 add read-only web UI and per-vendor readability prober
Web UI (internal/webui):
- dependency-free embedded HTML/JS served on the loopback listener
- /api/assets (provider/service/exposure/q filters) and /api/stats
- optional HTTP basic auth via UI_USER/UI_PASS

Prober (internal/probe):
- vendor-specific, unauthenticated, non-destructive checks:
  S3/R2/Spaces ListObjectsV2, GCS JSON object list, Azure ?comp=list,
  Firebase /.json?shallow=true, Supabase REST, and bare TCP connect for
  database endpoints (never reported readable, no auth attempted)
- exposure_detail records the factual reason (status code, region, tcp port)
- bounded random-sample queue instead of FIFO: one high-volume hostname
  template (mrap) no longer starves every other asset type
- schema: cloud_assets.exposure_detail + probed_at (idempotent ALTER)

Live: stable ~450MiB under 1GiB, no restarts, UI and probe states
(private/reachable/auth_required) flowing, 0 db errors.
2026-09-22 13:58:28 +02:00

93 lines
3.9 KiB
Markdown

# Architecture
## Data flow
```text
Certificate Transparency Logs
│
▼
CertSpotter internals fetch / verify / parse (third_party)
(in-process, no stdout/JSON/files)
│
▼
DNS identifiers
│
▼
FAST CLASSIFIER 0 allocs/op, no regexp/JSON/DNS/DB
┌────┴─────┐
direct cloud normal
resource hostname
│ │
│ ▼
│ DNS/CNAME enrichment bounded workers + TTL cache
│ │
│ ▼
│ classify CNAME targets
│ │
└─────┬──────┘
▼
dedup/enrichment TTL caches; DB uniqueness is final
│
▼
optional safe reachability HTTPS GET / only, OFF by default
▼
batch DB writer bounded queue, pgx batches
▼
PostgreSQL
```
## Stages and what is allowed
| Stage | May do | Must not do |
| --- | --- | --- |
| CertSpotter fetch/verify/parse | network, crypto, in-memory DER | persist certs, write files |
| Identifier callback | iterate `DNSNames`, call classifier | block, log per certificate, allocate needlessly |
| Classifier | byte/string comparisons, label scan | regexp, JSON, reflection, DNS, HTTP, DB, files, per-id logging |
| Direct queue | bound, non-blocking send | unbounded growth |
| DNS enrichment | DNS via worker pool, TTL cache, CNAME follow | unbounded goroutines, persistence |
| Dedup | memory-bounded TTL caches | unbounded maps |
| Probe (optional) | one unauthenticated vendor read endpoint (S3/GCS/Azure/Firebase list) or a bare TCP connect for databases | credentials, object downloads, mutation, DB queries, auth attempts |
| DB writer | batched `pgx.Batch` writes | one synchronous row per event |
| Checkpoints | JSON CT position in `ct_checkpoints`/`ct_sths` | certificate/issuer contents in DB |
## Concurrency and backpressure
* CertSpotter runs one monitor goroutine per CT log; it calls the identifier
sink concurrently.
* The sink never blocks: it classifies inline and does a non-blocking send to
either the direct queue or the DNS queue. Full queues increment
`queue_drops_total` / `queue_full_total`.
* Direct matches have a dedicated queue and worker pool, so speculative CNAME
enrichment can be shed first. DNS workers also skip work when the DB queue is
more than half full.
* The DB writer batches by size or `DB_FLUSH_INTERVAL` and flushes on
shutdown.
## Failure handling
* DNS and CT network errors are retried by CertSpotter with bounded exponential
backoff; ctcloud counts them (`ct_errors_total`) without logging per event.
* PostgreSQL writes retry a few times with backoff; persistent failure drops
items from the bounded DB queue rather than growing memory.
* The process uses `context` throughout and on `SIGTERM` stops intake, drains
the pipeline, flushes the DB batch and exits within `SHUTDOWN_TIMEOUT`.
## Metrics and health
`GET /metrics` exposes counters (`ct_certificates_total`,
`classifier_direct_matches_total`, `classifier_cname_matches_total`,
`dns_*`, `db_*`, `queue_*`) and gauges (`queue_depth{queue="…"}`,
`go_goroutines`, `process_system_bytes`). No hostnames or other high-cardinality
labels are ever used.
`GET /healthz` is a liveness check. `GET /readyz` verifies the CT monitor is
initialised and PostgreSQL is reachable.
## Why no files, pipes or subprocesses
Everything between certificate parsing and classification stays in memory. The
CertSpotter patch hands identifiers directly to Go code, removing serialization
(JSON), IPC (stdout/pipe), process boundaries (subprocess/hooks) and filesystem
I/O (certificate/state files). CT resume state is the only thing persisted, and
it lives in PostgreSQL.