pkg-proxy/docs/architecture.md

Architecture

This document describes the internal architecture of the git-pkgs proxy.

Overview

The proxy is a caching HTTP server that sits between package manager clients and upstream registries. It intercepts requests, checks a local cache, and either serves cached content or fetches from upstream.

┌──────────────────────────────────────────────────────────────────┐
│                          HTTP Server                              │
│  ┌──────────────────────────────────────────────────────────┐    │
│  │                     Router (Chi)                          │    │
│  │  /npm/*     -> NPMHandler      /health  -> healthHandler  │    │
│  │  /cargo/*   -> CargoHandler    /stats   -> statsHandler   │    │
│  │  /gem/*     -> GemHandler      /metrics -> prometheus     │    │
│  │  ...17 ecosystems              /api/*   -> APIHandler     │    │
│  │                                /ui/*    -> Web UI         │    │
│  └──────────────────────────────────────────────────────────┘    │
│         │                    │                    │               │
│         ▼                    ▼                    ▼               │
│  ┌───────────┐       ┌─────────────┐      ┌─────────────┐       │
│  │ Database  │       │   Storage   │      │   Upstream  │       │
│  │ SQLite or │       │ Filesystem  │      │  Registries │       │
│  │ Postgres  │       │  or S3      │      │  (Fetcher)  │       │
│  └───────────┘       └─────────────┘      └─────────────┘       │
└──────────────────────────────────────────────────────────────────┘

Request Flow

Metadata Request (npm example)

1. Client requests GET /npm/lodash
2. NPMHandler receives request
3. Handler fetches metadata from upstream registry.npmjs.org/lodash
4. Handler rewrites tarball URLs in metadata to point at proxy, and filters versions withheld by cooldown or the denylist (internal/metadata/npm)
5. Handler returns modified metadata to client

Metadata is not cached - always fetched fresh. This ensures clients see new versions immediately.

Artifact Download (npm example)

1. Client requests GET /npm/lodash/-/lodash-4.17.21.tgz

2. NPMHandler extracts package name and version from URL

3. Handler calls Proxy.GetOrFetchArtifact()

4. Proxy checks database for cached artifact:

   Cache Hit:

   • Look up artifact record in database
   • Open file from storage
   • Record hit (increment counter, update last_accessed_at)
   • Return reader to handler
   • Handler streams file to client, checking it against the recorded SHA-256 content_hash (the registry's own integrity was checked when the file was stored); a mismatch aborts the response and clears the record

   Cache Miss:

   • Resolve download URL using Resolver
   • Fetch artifact from upstream using Fetcher
   • Copy the body to Storage and, for a client GET without a Range header, to the client at once (storeWhileServing); the client's copy is checked against the upstream's digest as it is read, and a mismatch aborts it. A block-mode scanner, HEAD and range requests, and requests joining a fetch already running wait for the stored file.
   • Store artifact in Storage (returns size, hash)
   • Create/update database records (package, version, artifact)
   • Requests that waited open the stored file and stream it to their client

Artifact handlers honor single byte ranges when the storage reader supports efficient seeking (currently the local filesystem reader). They advertise Accept-Ranges: bytes, return 206 or 416 as appropriate, and respect If-Range. Malformed and multi-range requests fall back to the full response. Non-seekable storage readers do not advertise range support, and direct-storage redirects rely on the destination's capabilities. Range responses cannot verify the full artifact digest because they read only part of the artifact.

┌────────┐  GET /npm/lodash/-/lodash-4.17.21.tgz  ┌─────────────┐
│ Client │ ──────────────────────────────────────▶│ NPMHandler  │
└────────┘                                        └──────┬──────┘
                                                         │
                                                         ▼
                                               ┌─────────────────┐
                                               │ Proxy           │
                                               │ GetOrFetch      │
                                               └────────┬────────┘
                                                        │
                                    ┌───────────────────┼───────────────────┐
                                    │                   │                   │
                                    ▼                   ▼                   ▼
                             ┌───────────┐       ┌───────────┐       ┌───────────┐
                             │ Database  │       │  Storage  │       │ Upstream  │
                             │ (lookup)  │       │  (read)   │       │ (fetch)   │
                             └───────────┘       └───────────┘       └───────────┘

Package Structure

internal/database

SQLite or PostgreSQL database for cache metadata. SQLite uses modernc.org/sqlite (pure Go, no CGO). PostgreSQL uses lib/pq.

For SQLite, DB holds two pools: one connection for writes and a pool of read-only (query_only) connections for reads. Get, Select, Query and QueryRow go to the reader; Exec, transactions and everything else go to the writer, so a read inside a transaction sees the transaction's writes. A write sent to the reader fails instead of running. WAL mode lets the readers run while the writer commits. PostgreSQL and in-memory SQLite use one pool for both.

The schema is compatible with git-pkgs databases. The proxy adds the artifacts and vulnerabilities tables on top of the shared packages and versions tables, so both tools can point at the same database.

Tables:

packages (
    id          INTEGER PRIMARY KEY,  -- SERIAL on Postgres
    purl        TEXT NOT NULL,        -- unique, e.g. pkg:npm/lodash
    ecosystem   TEXT NOT NULL,
    name        TEXT NOT NULL,
    latest_version  TEXT,
    license         TEXT,
    description     TEXT,
    homepage        TEXT,
    repository_url  TEXT,
    registry_url    TEXT,
    supplier_name   TEXT,
    supplier_type   TEXT,
    source          TEXT,
    enriched_at     DATETIME,
    vulns_synced_at DATETIME,
    created_at      DATETIME,
    updated_at      DATETIME
)
-- indexes: purl (unique), (ecosystem, name)

versions (
    id           INTEGER PRIMARY KEY,
    purl         TEXT NOT NULL,       -- unique, e.g. pkg:npm/lodash@4.17.21
    package_purl TEXT NOT NULL,       -- FK to packages.purl
    license      TEXT,
    published_at DATETIME,
    integrity    TEXT,                -- subresource integrity hash
    yanked       INTEGER DEFAULT 0,  -- BOOLEAN on Postgres
    source       TEXT,
    enriched_at  DATETIME,
    created_at   DATETIME,
    updated_at   DATETIME
)
-- indexes: purl (unique), package_purl

artifacts (
    id             INTEGER PRIMARY KEY,
    version_purl   TEXT NOT NULL,
    filename       TEXT NOT NULL,
    upstream_url   TEXT NOT NULL,
    storage_path   TEXT,              -- null until cached
    content_hash   TEXT,              -- SHA-256
    size           INTEGER,           -- BIGINT on Postgres
    content_type   TEXT,
    fetched_at     DATETIME,
    hit_count      INTEGER DEFAULT 0, -- BIGINT on Postgres
    last_accessed_at DATETIME,
    created_at     DATETIME,
    updated_at     DATETIME
)
-- indexes: (version_purl, filename) unique, storage_path, last_accessed_at

pending_deletes (
    path          TEXT NOT NULL PRIMARY KEY, -- storage path no record points at
    queued_at     DATETIME NOT NULL
)

vulnerabilities (
    id            INTEGER PRIMARY KEY,
    vuln_id       TEXT NOT NULL,      -- e.g. CVE-2021-1234
    ecosystem     TEXT NOT NULL,
    package_name  TEXT NOT NULL,
    severity      TEXT,
    summary       TEXT,
    fixed_version TEXT,
    cvss_score    REAL,
    "references"  TEXT,               -- JSON array
    fetched_at    DATETIME,
    created_at    DATETIME,
    updated_at    DATETIME
)
-- indexes: (vuln_id, ecosystem, package_name) unique, (ecosystem, package_name)

metadata_cache (
    id            INTEGER PRIMARY KEY,
    ecosystem     TEXT NOT NULL,
    name          TEXT NOT NULL,
    storage_path  TEXT NOT NULL,
    etag          TEXT,
    content_type  TEXT,
    content_encoding TEXT,           -- replayed on serve so signed bytes stay verbatim
    size          INTEGER,           -- BIGINT on Postgres
    fetched_at    DATETIME,
    created_at    DATETIME,
    updated_at    DATETIME
)
-- indexes: (ecosystem, name) unique

metadata_dates (
    ecosystem     TEXT NOT NULL,
    proxy_url     TEXT NOT NULL,      -- each proxy URL rewrites to its own download URLs
    name          TEXT NOT NULL,
    etag          TEXT NOT NULL,      -- the rewritten representation being dated
    last_modified INTEGER NOT NULL,   -- Unix seconds; BIGINT on Postgres
    PRIMARY KEY (ecosystem, proxy_url, name)
)

metadata_dates holds the Last-Modified served with rewritten metadata, such as npm packuments, Composer documents, PyPI simple pages and Helm indexes. A package's date is the second it started serving its current representation, and a new representation is only dated in a later second. Keeping it in the database means replicas sharing one serve the same date, and never give one representation's date to another, so If-Modified-Since cannot return 304 for a body the client does not have.

On PostgreSQL, INTEGER PRIMARY KEY becomes SERIAL, DATETIME becomes TIMESTAMP, INTEGER DEFAULT 0 booleans become BOOLEAN DEFAULT FALSE, and size/count columns use BIGINT.

The MigrateSchema() function handles backward compatibility with older git-pkgs databases by running named migrations that add missing columns and tables. See migrations.md for how to add new schema changes.

Key operations:

• GetPackageByPURL() - Look up package by PURL
• GetVersionByPURL() - Look up version by PURL
• GetArtifact() - Look up artifact by version + filename
• UpsertPackage/Version/Artifact() - Insert or update records
• RecordArtifactHit() - Increment hit counter, update access time
• GetLeastRecentlyUsedArtifacts() - For cache eviction
• SearchPackages() - Full-text search across cached packages

internal/storage

File storage abstraction. Current implementation uses local filesystem.

Interface:

type Storage interface {
    Store(ctx, path, reader) (size, hash, error)
    Open(ctx, path) (io.ReadCloser, error)
    Exists(ctx, path) (bool, error)
    Delete(ctx, path) error
    Size(ctx, path) (int64, error)
    UsedSpace(ctx) (int64, error)
}

Filesystem implementation:

• Stores files in nested directories: {ecosystem}/{name}/{version}/{fetch id}/{filename}, a new id per fetch, so fetches of one artifact never overwrite or delete each other's object (artifacts cached before this sit at {ecosystem}/{name}/{version}/{filename})
• Atomic writes using temp file + rename
• Computes SHA256 hash during write
• Removes a fetch's directory once its object is deleted

Path structure:

cache/artifacts/
├── npm/
│   ├── lodash/
│   │   └── 4.17.21/
│   │       └── 3f9a0c1d2e4b5a67/
│   │           └── lodash-4.17.21.tgz
│   └── @babel/
│       └── core/
│           └── 7.23.0/
│               └── 8c2e41f09a7d3b15/
│                   └── core-7.23.0.tgz
└── cargo/
    └── serde/
        └── 1.0.193/
            └── d05b7e9c14a2f863/
                └── serde-1.0.193.crate

internal/upstream

Fetches artifacts from upstream registries.

Fetcher:

• HTTP client with configurable timeout (5 min default for large artifacts)
• Exponential backoff retry on 429 (rate limit) and 5xx errors
• Returns streaming reader (doesn't load into memory)
• Configurable user-agent
• Shares an authentication-aware transport with metadata requests so URL-scoped credentials apply consistently
• Discovers and caches scoped OCI Bearer tokens from registry challenges

Resolver:

• Determines download URL for a package/version
• Handles ecosystem-specific URL patterns:
  • npm: https://registry.npmjs.org/{name}/-/{shortname}-{version}.tgz
  • cargo: https://static.crates.io/crates/{name}/{name}-{version}.crate
  • etc.

internal/handler

HTTP protocol handlers for each registry type.

Proxy (shared):

• GetOrFetchArtifact() - Main cache logic
• Coordinates database, storage, and fetcher
• Handles cache hit/miss flow

NPMHandler:

• handlePackageMetadata() - Proxy + rewrite metadata
• handleDownload() - Serve cached artifact
• Fetches, caches and serves packuments, and caches their rewrites; the rewrite itself is internal/metadata/npm

CargoHandler:

• handleConfig() - Return registry config
• handleIndex() - Proxy sparse index
• handleDownload() - Serve cached crate

SwiftHandler:

• Proxies the Swift Package Registry v1 read endpoints
• Rewrites release URLs and caches source archives
• With cache_metadata, caches release metadata, release lists and manifests (with their Link header) and serves them when the upstream fails

GitHandler:

• Serves /git/{name}/{repo} from internal/gitcache, read-only over smart HTTP
• Validates repository paths and the repos allowlist, and maps cache errors to statuses
• Serves the Git LFS API under {repo}.git/info/lfs/: forwards download batches upstream and points each object at objects/{oid}/{size}, which is served from the artifact cache (ecosystem git, name {name}/{repo}, version the object ID); a miss uses the download URL kept from the batch or asks the upstream batch API for a fresh one, recorded without its signed query, copies the object from another repository when its content is already stored, and checks the object against its sha256 ID; a batch the upstream can't answer is answered from the cache

internal/metadata/npm

The npm metadata rewrite, separated from the handler so it can run where the proxy's infrastructure can't, such as WASI. It takes a packument's bytes and everything else as explicit inputs (proxy URL, cooldown policy, the time cooldown is evaluated at, denied versions, pushed versions) and returns the rewritten document with the versions it kept. It filters versions by cooldown and the denylist, moves latest off a withheld version, lists pushed versions in place of upstream ones, and points each tarball URL at the proxy, copying everything else byte for byte. It does no I/O and never reads the clock; callbacks report what it did, which the handler logs. It imports no database, storage, metrics, HTTP or logging packages, which a test checks.

internal/jsonscan

Reads and rewrites members of a JSON document in place, by byte offsets, without decoding the whole document. Shared by the npm and Composer metadata rewrites.

internal/gitcache

Local bare clones of upstream Git repositories, outside the artifact cache and the database, at {git.cache_dir}/{name}-{URL hash}/{repo}.git. The hash of the upstream's URL means repointing an upstream starts a fresh cache; at startup only the configured upstreams' directories are loaded, and those of their old URLs removed.

• Each request for refs checks the upstream's branches and tags with protocol v2 ls-refs and fetches only when they changed; concurrent requests share one refresh
• Clients are served by git upload-pack --stateless-rpc
• Spawned git reaches the upstream only through a loopback relay that forwards with the proxy's upstream client, so address checks, upstream.auth and the access log apply
• Commits a client asks for by ID that are on no cached branch or tag are fetched and kept under hidden refs/proxy/wanted/{time}/{id} refs until a branch or tag includes them or for 7 days; off for upstreams with credentials
• Evicts least recently used repositories after each fetch once git.max_size is passed, and drops a repository that can no longer be refreshed within git.max_repo_size so it is cloned again; never one in use. A commit fetched by ID that does not fit is refused instead
• Every fetched pack is kept as a pack (transfer.unpackLimit=1), so a transfer cut off at the size limit leaves nothing behind

internal/server

HTTP server setup, web UI, and API handlers.

• Creates and wires together all components
• Mounts protocol handlers at ecosystem-specific paths
• Middleware: request ID, real IP, logging, panic recovery, active request tracking
• Web UI under /ui: dashboard, package browser, source browser, version comparison
• Templates are embedded in the binary via //go:embed
• Enrichment API for package metadata, vulnerability scanning, and outdated detection
• Health, stats, and Prometheus metrics endpoints. /health runs an active write → size-check → read → verify → delete probe against the storage backend and returns a structured JSON response (HealthResponse) with "ok" / "error" status per subsystem. Probe results are cached (default 30 s, configurable via health.storage_probe_interval) to avoid overwhelming remote backends. The response also carries a circuit_breakers map reporting each upstream's artifact-fetch breaker as "open" or "closed", keyed by the host fetched from (or an opaque placeholder where the fetch URL has no host to read); the same state is published as the proxy_circuit_breaker_state gauge on each /metrics scrape. An open breaker leaves the overall status "ok" — it describes an upstream, not this proxy.

internal/memlimit

Sets Go's soft memory limit to 80% of the cgroup memory limit at startup on Linux, unless GOMEMLIMIT is set. See "Memory limit" in the configuration docs.

internal/metrics

Prometheus metrics for cache performance, upstream latency, storage operations, and active requests. See the Monitoring section of the README for the full metric list.

Cooldown

Version age filtering for supply chain attack mitigation, provided by github.com/git-pkgs/cooldown. Configurable at global, ecosystem, and per-package levels. Supported by the npm, PyPI, pub.dev, Composer, Cargo, NuGet, Conda, RubyGems, Hex, Go, Maven and Helm handlers.

internal/enrichment

Package metadata enrichment, built on github.com/git-pkgs/enrichment. Fetches license, description, homepage, repository URL, and vulnerability data from upstream registries. Powers the /api/ endpoints and the web UI's package detail pages.

internal/mirror

Selective package mirroring for pre-populating the proxy cache. Supports multiple input sources: individual PURLs (versioned or unversioned), CycloneDX/SPDX SBOM files, and full registry enumeration. Uses a bounded worker pool backed by errgroup to download artifacts in parallel, reusing handler.Proxy.GetOrFetchArtifact() for the actual fetch-and-cache work. Container images (pkg:docker, pkg:oci) go through handler.ContainerHandler.MirrorImage() instead, which walks index, platform manifests and blobs and stores them under the keys a pull reads.

The package also provides a MetadataCache for storing raw upstream metadata blobs so the proxy can serve metadata responses offline. The JobStore manages async mirror jobs exposed via the /api/mirror endpoints.

internal/config

Configuration loading.

• Supports YAML and JSON files
• Environment variable overrides (PROXY_ prefix)
• Command line flag overrides
• Validation

Extending the Proxy

Adding a New Registry

1. Add URL resolution in upstream/resolver.go
2. Create handler in handler/newregistry.go
3. Mount in server/server.go
4. Add tests

Adding a New Storage Backend

1. Implement storage.Storage interface
2. Add configuration options in config/config.go
3. Add initialization in server/server.go

Cache Eviction

The database tracks hit_count and last_accessed_at for LRU eviction. Query with:

db.GetLeastRecentlyUsedArtifacts(limit)

Eviction can be implemented as:

1. Background goroutine checking GetTotalCacheSize()
2. When over limit, get LRU artifacts
3. Delete from storage and clear database records

An object a record stops pointing at, because a refetch replaced it or its entry was discarded, is not deleted at once: a request that read the record may still be opening it. Its path goes into pending_deletes, and a background loop deletes it after a grace period of at least an hour, or direct_serve_ttl if longer, so signed URLs to it stay valid. This runs whether or not max_size is set.

Design Decisions

Why SQLite?

• Simple deployment (single file)
• No external dependencies
• Good performance for this workload
• Pure Go driver available (no CGO)

Why rewrite metadata URLs?

• Ensures clients fetch artifacts through proxy
• Alternative: Let clients fetch directly, miss cache opportunity

Why not cache metadata (by default)?

• Simplicity - no invalidation logic needed
• Fresh data - new versions visible immediately
• Metadata is small, upstream fetch is fast
• Set cache_metadata: true or use the mirror command to enable metadata caching for offline use via the metadata_cache table
• OCI manifests and tag lists are exceptions: they are cached automatically so previously fetched images remain pullable and tag resolution works when the registry or token service is unavailable

Why stream artifacts?

• Memory efficient - don't load large files into RAM
• Better latency - start sending while still receiving

Why atomic writes?

• Prevents serving partial files
• Safe concurrent access
• Clean recovery from crashes