QuotaProvider grows a read half: CheckRead(org, bytes) and RecordEgress(org, bytes). CheckRead is enforced on /s/* only — a public share link is the sole unauthenticated door to stored bytes, so it is the only egress a plan can cap. The sync proxy and viewer merely RecordEgress: refusing a device mid-cycle surfaces as ErrForbidden, which the syncer reads as "access is gone — pause and touch nothing", and sync must never break over a bill. UnlimitedQuota stays the OSS default. countingWriter bills what actually reached the client rather than a size claimed before the write. bdrive init warns past 1 GiB or 20k files and says how to narrow scope; syncer.Measure sizes that through the real Filter and the one walkFolder predicate. starterIgnore gains video/archive/disk-image globs and Library/ — every version is kept forever, so a big binary committed once is paid for forever on every device. deploy: a Nearline-at-30-days lifecycle rule and the arithmetic for why it stops there. Coldline and Archive only pay off below roughly one read per month, and a first sync pulls every historical blob rather than just the current tree, so blob read rate tracks device onboarding. docs/launch-plan.md said Cloud was waitlist-only and framed Product Hunt as an OSS launch whose goal was not signups or revenue; both are stale. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
6.7 KiB
Deploy BearDrive to Google Cloud (Cloud Run)
Single-instance hub on Cloud Run, metadata in Cloud SQL Postgres, blobs/journals in a GCS bucket. Matches Phase 0 of the managed PRD.
The current build is single-process: Cloud Run is pinned to
max-instances=1because the in-memory caches assume one writer. Do not raise it until the "stateless app" work (PRD §5.1) lands.
Architecture
browser / bdrive CLI
│ https
┌─────▼─────┐ metadata ┌──────────────┐
│ Cloud Run │◀────────────▶│ Cloud SQL │ (accounts, orgs,
│ bdrive │ unix socket │ Postgres │ projects, invites…)
│ (1 inst.) │ └──────────────┘
└─────┬─────┘
│ ADC (runtime SA)
┌─────▼─────┐
│ GCS │ blobs/ + journal/ (file content + sync log)
└───────────┘
Prerequisites
gcloudinstalled and logged in (gcloud auth login).- A billing account id (
gcloud billing accounts list) if the script creates the project. - Values:
PROJECT_ID,ADMIN_EMAIL,ADMIN_DOMAIN(the rest have defaults).
Run it
From the repo root:
PROJECT_ID=beardrive-prod \
BILLING_ACCOUNT=0X0X0X-0X0X0X-0X0X0X \
ADMIN_EMAIL=you@runbear.io \
ADMIN_DOMAIN=runbear.io \
REGION=us-central1 \
bash example/deploy/gcp-cloudrun.sh
The script: creates/links the project → enables APIs → creates the GCS bucket
and Cloud SQL instance → generates a DB password (stored in Secret Manager) →
writes the hub config to a secret → builds the image from the repo Dockerfile
via Cloud Build → deploys Cloud Run with the Cloud SQL socket, the config
secret mounted at /config/config.json, and a dedicated runtime service
account granted GCS + Cloud SQL access. It prints the service URL.
First-run: bootstrap the admin, then lock down
The hub ships invite-only by default, but a brand-new hub has no accounts,
so the deploy config temporarily allows domain-gated self-signup
(allowed_domains: [ADMIN_DOMAIN]). Steps:
- Open the printed URL → Sign up as
ADMIN_EMAIL(must be onADMIN_DOMAIN). The account is active immediately and is a hub admin. - Create your org/projects and invite teammates from the UI.
- Tighten to invite-only: edit the config secret to
"allow_signup": falseand redeploy:gcloud secrets versions access latest --secret bdrive-config > /tmp/c.json # …set "allow_signup": false … gcloud secrets versions add bdrive-config --data-file=/tmp/c.json gcloud run services update bdrive --region "$REGION" # picks up latest secret
Rough cost
- Cloud Run: scales to ~zero when idle (min-instances=1 keeps one warm;
set
--min-instances 0to save more, at the cost of cold-start journal folding on first hit). ~$5–15/mo warm. - Cloud SQL
db-f1-micro: ~$8–15/mo (smallest shared-core tier). - GCS: pay per GB stored + egress. Cheap for text.
gcp-cloudrun.shinstalls a Nearline-at-30-days lifecycle rule by default (LIFECYCLE=0skips it) — see below for why it stops there.
Storage tiering
Every version of every file is retained forever, so stored bytes only ever grow while per-seat revenue stays flat. Aging objects down is the lever.
Prices below are us-central1 regional, verified 2026-08-03 — re-check before relying on them, GCP moves them.
| Class | Storage $/GB/mo | Retrieval $/GB | Min duration |
|---|---|---|---|
| Standard | 0.020 | — | none |
| Nearline | 0.010 | 0.01 | 30 days |
| Coldline | 0.004 | 0.02 | 90 days |
| Archive | 0.0012 | 0.05 | 365 days |
Archive really is ~6% of Standard, and GCS serves every class at the same
millisecond latency — there is no restore job to wait on. The catch is the
retrieval fee, so the break-even is entirely about how often a given object
is read. Writing r for reads per GB per month:
- Nearline beats Standard while
r < 1.0/mo - Coldline beats Nearline while
r < 0.6/mo - Archive beats Coldline while
r < 0.09/mo(about once a year)
This is why the script stops at Nearline. The tempting assumption is that
old blobs are cold because old versions are rarely opened. That is not true
here: a device syncing a project for the first time downloads a blob for
every put op in every peer journal — the entire history, not just the
current file tree (internal/syncer pull). Measured on a 10-version file
whose working tree is 1 KB, a fresh device pulls 10 KB.
So the read rate on old blobs tracks how often anyone adds a device, not
how often anyone opens an old version. A team that adds or replaces roughly
one device a month drives r ≈ 1, which makes Coldline a wash and Archive a
straight bill increase.
Two consequences worth acting on, in this order:
- The real lever is not the lifecycle policy. Making a first sync fetch
only current-state blobs (history stays available on demand through the
existing
/blob?sha=route) cuts onboarding egress from "all history" to "the working tree" and makes old blobs genuinely cold — which is what makes Coldline and Archive safe to turn on afterwards. Until then the ladder is priced against a read pattern the sync engine does not have. - Egress scales with devices × total history, not with change volume,
and every byte is relayed: blob reads have no presigned path (only
remote.PutSignerexists — uploads can go direct to storage, downloads cannot), so they stream GCS → Cloud Run → client. Same-region GCS→Cloud Run transfer is free, so this is one egress charge, not two — but it does occupy the singlemax-instances=1container for the whole transfer.
The lifecycle rule is applied bucket-wide rather than to blobs/ alone
because journals live under the same per-project prefixes and a GCS
lifecycle matchesPrefix cannot express */blobs/. That is safe: a peer
journal is re-fetched only when the listing shows it grew, and one that grew
was just rewritten, so it is Standard again.
Notes / limits (single-instance build)
max-instances=1is required. Metadata correctness depends on one writer.BDRIVE_HOME=/tmpis ephemeral on Cloud Run → the server's own device id regenerates on cold start (cosmetic in history). Mount a volume later to persist it.- Large downloads/sync stream through Cloud Run (bounded by the request timeout, up to 60 min). Uploads go direct to storage when the backend can presign. See PRD §5.2 for offloading these at scale.
- Put a custom domain on the service via
gcloud run domain-mappings(gives managed TLS).