# Roadmap: Producer HTTP Upload Package ## Purpose Add a second public producer-facing package that lets Go producer applications build or validate a compliant source bundle, package it as a gzip-compressed tar archive, and submit it to the HTTP upload API with safe retry support. The current public producer package, `pkg/bundle`, owns source manifest semantics, digest calculation, path validation, local manifest building, local bundle writing, and local bundle validation. The new package must build on that contract instead of reimplementing it. This roadmap also adds server-side producer idempotency keys to the HTTP upload API. Idempotency is required for the producer upload package's retry behavior: the client can safely retry an upload with the same key, and the server can collapse duplicate accepted uploads into the original run. ## Goals - Make the common Go producer workflow small and hard to misuse. - Reuse `pkg/bundle` for manifest generation, path normalization, SHA-256 calculation, digest calculation, and local validation. - Create upload archives that match the server's source bundle archive contract. - Add server-side idempotency records scoped to the authenticated pipeline. - Send idempotency keys from the public upload package by default. - Handle bearer authentication without logging or returning token values. - Parse successful, duplicate, conflict, and error responses into typed producer-side results. - Retry safely using idempotency keys and bounded backoff. - Keep the package dependency-light and usable from ordinary Go producer applications. ## Non-Goals - Do not expose `internal/app`, `internal/ingest`, storage backends, server config, or destination state types through the public package. - Do not add durable idempotency storage, durable client queues, background workers, or database-backed retry processing. - Do not add zstd, multipart upload, resumable upload, or non-tar archive formats. - Do not require producers to know distributor pipeline ids; server-side token mapping remains authoritative. - Do not make the package a replacement for the existing CLI or server API documentation. ## Implementation Sequence Implement this feature in three stages: 1. Server-side HTTP idempotency keys. 2. Public `pkg/upload` client package. 3. Current-behavior documentation and examples. Server idempotency should land first so the public upload package can rely on the final retry contract from its first release. ## Stage 1: Server Idempotency Keys Goal: Extend `distributor serve` so `POST /upload` can safely accept retried producer uploads without creating duplicate accepted runs. HTTP contract: - Producers may send `Idempotency-Key: ` with `POST /upload`. - The public upload package must always send this header. - Raw HTTP clients may omit it; omitted keys preserve current behavior. - Keys are scoped to the authenticated pipeline selected by bearer token. - Valid keys are non-empty ASCII strings up to 128 bytes using letters, digits, `.`, `_`, `-`, and `:`. - Invalid keys return `400`. Server behavior: - After archive staging and source bundle validation succeeds, record the idempotency key with the accepted run id and the normalized source manifest identity. - If the same pipeline receives the same key and the staged upload has the same normalized source manifest identity, return the original accepted response instead of enqueueing another run. - If the same pipeline receives the same key and the staged upload has a different normalized source manifest identity, return `409 Conflict`. - If the same key is already being processed concurrently for the same pipeline before a manifest identity is available, return a retryable conflict response without accepting a new run. - Idempotency records are memory-only and expire with the existing HTTP upload retention window. - Server restart clears idempotency records, matching the current memory-only status and queue behavior. Manifest identity: - Compare normalized source manifest semantics, not raw archive bytes. - At minimum, compare manifest schema version, id, created timestamp, bundle digest, and ordered file records. - Different tar metadata or gzip encoding for the same source bundle should not create a conflict. Tests: - `go test ./internal/app ./internal/ingest` - Same token, same key, same staged bundle returns the original run id and does not enqueue a second run. - Same token, same key, different staged bundle returns `409`. - Same key under different authenticated pipelines does not conflict. - Missing idempotency key preserves existing raw HTTP behavior. - Invalid key syntax returns `400`. - Idempotency records expire with completed run status retention. - Secret tokens and idempotency keys are not logged in errors beyond the key value itself where required for diagnostics. Completion criteria: - The HTTP API has an implemented, tested idempotency contract. - Existing clients without `Idempotency-Key` continue to work. - Duplicate idempotent uploads cannot create duplicate accepted runs. ## Stage 2: Public `pkg/upload` Client Goal: Add a new public `pkg/upload` package that submits compliant bundles to `distributor serve` using bearer authentication and idempotency keys. Package name: - Use `pkg/upload`. - Rationale: `pkg/bundle` owns bundle construction and validation; `pkg/upload` owns submission to the distributor HTTP upload API. API shape: Use options-struct APIs rather than one large positional function. Initial APIs must cover two producer workflows: - upload an existing local bundle root; - build a temporary bundle from explicit producer files and upload it. Representative API shape: ```go package upload type Client struct { // unexported fields } type ClientOptions struct { Endpoint string Token string HTTPClient *http.Client Retry RetryOptions } type RetryOptions struct { MaxAttempts int BaseDelay time.Duration MaxDelay time.Duration } type UploadBundleOptions struct { Root string Validate bool IdempotencyKey string } type UploadFilesOptions struct { ID string Created time.Time Files []bundle.BundleFile Validate bool TempDir string IdempotencyKey string } type Result struct { RunID string Status string } func NewClient(opts ClientOptions) (*Client, error) func (c *Client) UploadBundle(ctx context.Context, opts UploadBundleOptions) (Result, error) func (c *Client) UploadFiles(ctx context.Context, opts UploadFilesOptions) (Result, error) func (c *Client) Status(ctx context.Context, runID string) (RunStatus, error) ``` Required API semantics: - `Endpoint` is a distributor server base URL. The client derives `/upload` and `/runs/` internally. - `Token` is required and is sent as `Authorization: Bearer `. - `HTTPClient` is optional; when omitted, use a client with conservative timeouts. - `UploadBundle` reads and packages an existing local source bundle. - `UploadFiles` creates a temporary complete bundle through `pkg/bundle`, then packages and uploads it. - `Status` is optional for callers and never required by `UploadBundle` or `UploadFiles`. - `Result` represents upload admission, not final publication success. Idempotency key behavior: - The client must send `Idempotency-Key` on every upload. - If the caller supplies `IdempotencyKey`, use it. - If omitted, generate a random 128-bit lowercase hex key once for that upload operation and reuse it for all retries from that call. - Generated keys are not stable across process restarts or separate calls. - Producers that need cross-process retry safety must supply their own stable key. - Validate caller-supplied keys before making a request. Bundle and archive behavior: - `UploadBundle` loads `manifest.json` from the bundle root and validates the local bundle by default. - `UploadBundle` includes `manifest.json` and every manifest-listed file in the tar.gz archive, and does not include unlisted files. - `UploadFiles` requires a non-empty bundle id and non-empty file list. - `UploadFiles` uses `pkg/bundle.WriteBundle` or equivalent public bundle APIs in a temporary directory and preserves explicit file order. - Zero `Created` follows `pkg/bundle` defaulting behavior. - Validation is enabled by default and may be explicitly disabled only for callers that already performed equivalent validation. - Tar entry names are slash-separated bundle-relative paths. - The package must not write into producer source directories. Archive and retry strategy: - Create a replayable upload body for each upload operation. - The implementation may either create a temporary `.tar.gz` file or regenerate the tar.gz body from the validated staged bundle for each attempt. - Clean up all temporary bundles and archive files created by the package. - Use `Content-Type: application/gzip`. Retry policy: - Defaults should be safe and modest, for example three total attempts with bounded exponential backoff. - Retry `503 Service Unavailable` because the upload was not accepted. - Retry temporary network errors and ambiguous mid-upload failures using the same idempotency key and replayable body. - Do not retry `400`, `401`, `409`, `413`, or `415`. - Do not retry after `202 Accepted`. - Respect caller context cancellation before waiting and before each retry. - Redact the bearer token from all errors. HTTP response handling: - Parse `202 Accepted` responses into `Result`. - Parse JSON error bodies where available. - Include HTTP status codes and response messages in typed errors. - Treat duplicate idempotent `202` responses the same as first acceptance. - Treat `409 Conflict` as an idempotency conflict error. - Close response bodies on every attempt. Tests: - `go test ./pkg/bundle ./pkg/upload` - Client construction with valid and invalid base endpoints. - Missing token rejection and token redaction in errors. - Caller-supplied and generated idempotency keys. - Uploading an existing valid bundle root. - Building and uploading from `bundle.BundleFile` values. - Local validation failures before any HTTP request. - Tar.gz entry names, manifest inclusion, and exclusion of unlisted files. - `202`, `400`, `401`, `409`, `413`, `415`, `503`, non-JSON errors, and unexpected status response parsing. - Safe retry with the same idempotency key for `503` and retryable network failures. - No retry for non-retryable statuses. - Context cancellation during retry backoff. - Custom `*http.Client` behavior through `httptest.Server`. Completion criteria: - Producer applications can build or validate a bundle and upload it with one package. - All uploads include idempotency keys. - Retry behavior is safe under the server idempotency contract. ## Stage 3: Documentation And Examples Goal: Document the implemented producer upload package and idempotency behavior only after the server and public package exist. Current-behavior documentation updates: - `README.md`: mention the new producer upload package briefly. - `docs/integrations/source-bundle.md`: link from producer APIs to upload helpers. - `docs/integrations/http-upload.md`: document `Idempotency-Key` and add a short Go producer helper section. - `docs/operations.md`: add a concise producer-side example if useful. - `docs/internal/app.md`: document server-side idempotency record behavior. - `docs/policy/development.md`: document the `pkg/upload` boundary and test expectations. Tests and checks: ```sh go test ./... rg -n "pkg/upload|Idempotency-Key|UploadBundle|UploadFiles" README.md docs examples ``` Completion criteria: - Current docs describe implemented behavior. - Future-only behavior remains under `docs/roadmap/`. ## Deferred Work - Durable idempotency records across server restarts. - Producer-supplied idempotency keys integrated with a database-backed queue. - `UploadAndWait` or long-polling helpers. - Run cancellation, retry, or listing endpoints. - Zstandard-compressed tar archives. - Multipart, resumable, or streaming object upload support. - URL-token authentication for constrained clients.