@interop/was-client 0.97.1 → 0.99.0

package/dist/edv/EdvCodec.js

@@ -38,13 +38,15 @@
  *   is encrypted as one document -- stored as a legible UTF-8 string for a
  *   text-family type (else base64) -- with the plaintext content type and the
  *   encoding carried in the document `meta`. A blob over `maxBlobBytes`, and
- *   any `ReadableStream` (whose size is unknown), is auto-routed by `add()` to
- *   the chunked-stream path instead: `encode` returns a multi-request plan the
- *   write path executes, storing one document plus its chunk resources over a
- *   `WasTransport` supplied by the injected `CodecTransportFactory`. Reads
- *   reassemble transparently, so `get()` returns the same `Blob` either way,
- *   and `decodeStream` hands a chunked document's decrypt stream back
- *   unbuffered.
+ *   any `ReadableStream` (whose size is unknown), is auto-routed by `add()` and
+ *   `put()` alike to the chunked-stream path instead: `encode` returns a
+ *   multi-request plan the write path executes, storing one document plus its
+ *   chunk resources over a `WasTransport` supplied by the injected
+ *   `CodecTransportFactory`. A `put()` over a chunked document is a plan too,
+ *   whatever its payload: the prior document's chunks that the new one does
+ *   not cover are reaped after the write. Reads reassemble transparently, so
+ *   `get()` returns the same `Blob` either way, and `decodeStream` hands a
+ *   chunked document's decrypt stream back unbuffered.
  * - **Enforced sequence (conditional writes).** The codec sets
  *   `conditionalWrites`, so the write path pre-reads the current envelope and
  *   hands it to `encode`: an update advances `sequence` from its prior value and
@@ -102,6 +104,13 @@ export const DEFAULT_MAX_BLOB_BYTES = 512 * 1024;
  * decrypted document self-describing alongside `'utf-8'` and `'base64'`.
  */
 const CHUNKED_ENCODING = 'chunked';
+/**
+ * How many stale chunks a reap deletes at once. Each delete is one signed
+ * request that is independent of the others (pinned and 404-tolerant, in no
+ * required order), so a shrunk document's leftovers go a few at a time rather
+ * than one awaited round trip per chunk.
+ */
+const REAP_CONCURRENCY = 8;
 /**
  * A shared strict UTF-8 decoder used to test whether a non-JSON payload is
  * valid UTF-8 (so it can be stored legibly as text rather than base64).
@@ -452,6 +461,41 @@ export class EdvCodec {
                 ? undefined
                 : (await this.#edv.generateId()));
         const parts = await this.#toDocument(data, contentType, docId);
+        // When the write path pre-read a current envelope, this is an update:
+        // `sequence` advances from the prior value (`encrypt({ update: true })`
+        // increments it), the write is pinned to the server's current ETag with
+        // `If-Match`, and the write learns whether the prior document's bytes
+        // live in chunk resources, so they can be reaped once this write lands.
+        // Only the sealed `meta.encoding` says so for certain, and reading it
+        // means opening the envelope (decrypting it, with its binding verified
+        // against this id, as a read would): a key unwrap per update. The
+        // cleartext `stream` member a chunked write leaves on the envelope is the
+        // cue to pay for that; an envelope without one is taken as inline and is
+        // not opened. The cue is server-controlled, but nothing rests on it: a
+        // server that strips it can only leave the prior chunks stranded (stored
+        // and counted against quota, unreachable, removed with the resource), and
+        // one that adds it costs this write a decrypt that finds an inline
+        // encoding. With no prior envelope this is a fresh insert (`sequence: 0`)
+        // guarded by `If-None-Match: *` (create-if-absent), so a concurrent first
+        // writer cannot be clobbered.
+        let prior = null;
+        if (current) {
+            const read = await readJsonData(current);
+            this.#assertEnvelope(read, 'update');
+            const chunked = read.stream !== undefined &&
+                (await this.#openEnvelope({ doc: read, expectedId: id })).meta
+                    ?.encoding === CHUNKED_ENCODING;
+            // The caller's own compare-and-swap baseline wins when they named one:
+            // the write path has already checked it against the pre-read, so the
+            // two agree by the time this runs.
+            const etag = precondition?.ifMatch ?? readEtag(current);
+            prior = {
+                sequence: read.sequence,
+                ...(read.indexed !== undefined && { indexed: read.indexed }),
+                ...(etag !== undefined && { etag }),
+                chunked
+            };
+        }
         if (parts.kind === 'chunked') {
             if (docId === undefined) {
                 // A stream is chunked whatever its size, so only a sized payload can
@@ -473,21 +517,11 @@ export class EdvCodec {
             return this.#chunkedWrite({
                 id: docId,
                 stream: parts.stream,
-                meta: parts.meta
+                meta: parts.meta,
+                prior
             });
         }
         const { content, meta } = parts;
-        // When the write path pre-read a current envelope, advance `sequence` from
-        // its prior value (`encrypt({ update: true })` increments it) and pin the
-        // write to the server's current ETag with `If-Match`. With no prior envelope
-        // this is a fresh insert (`sequence: 0`) guarded by `If-None-Match: *`
-        // (create-if-absent), so a concurrent first writer cannot be clobbered.
-        let priorDoc = null;
-        if (current) {
-            const read = await readJsonData(current);
-            this.#assertEnvelope(read, 'update');
-            priorDoc = read;
-        }
         const { documentCipher } = this.#edv;
         const encrypted = await documentCipher.encrypt({
             doc: {
@@ -497,14 +531,14 @@ export class EdvCodec {
                 // the plaintext content type and the inline-encoding discriminator,
                 // taken fresh from this write (the new type/encoding wins on update).
                 meta,
-                ...(priorDoc && { sequence: priorDoc.sequence }),
+                ...(prior && { sequence: prior.sequence }),
                 // Carry the prior envelope's blinded `indexed` entries into the
                 // update. With no blinding key edv-client stores `doc.indexed`
                 // verbatim; omitting it would replace the entries with `[]` and
                 // silently drop the record from `find()`. With a blinding key the
                 // entries are recomputed from the schema and this copy is the prior
                 // state that recomputation updates.
-                ...(priorDoc?.indexed !== undefined && { indexed: priorDoc.indexed })
+                ...(prior?.indexed !== undefined && { indexed: prior.indexed })
             },
             // Bind an AEAD-authenticated `was` parameter into the JWE protected
             // header: the scheme version, the resource id when known at encrypt time
@@ -521,7 +555,7 @@ export class EdvCodec {
                 ...(docId !== undefined && { resource: docId }),
                 hmac: this.#writeBlindingKey()
             }),
-            update: priorDoc !== null
+            update: prior !== null
         });
         if (docId === undefined) {
             // Encrypt-then-stamp: the id lives in the cleartext envelope, outside the
@@ -530,6 +564,12 @@ export class EdvCodec {
             docId = await documentCipher.deriveId({ jwe: encrypted.jwe });
             encrypted.id = docId;
         }
+        if (prior?.chunked) {
+            // The envelope replaces a document whose bytes live in chunk resources,
+            // which the inline document leaves behind: a plan writes it and reaps
+            // them, since one request cannot do both.
+            return this.#replacingWrite({ id: docId, encrypted, meta, prior });
+        }
         // Serialized on first read of `body` and kept for any later read, so the
         // wire bytes are built once per write and only when someone asks for them.
         let bytes;
@@ -564,9 +604,7 @@ export class EdvCodec {
             // (create-only-if-absent), and the server enforces it, so a masked-404
             // pre-read cannot silently overwrite an existing document.
             ...(precondition ??
-                (priorDoc
-                    ? { ifMatch: readEtag(current ?? null) }
-                    : { ifNoneMatch: true })),
+                (prior ? { ifMatch: prior.etag } : { ifNoneMatch: true })),
             // Stamp the key epoch this write encrypted under (the `currentEpoch`), so
             // the server records it and a reader can pick the epoch key.
             epoch: this.#writeEpoch
@@ -579,9 +617,11 @@ export class EdvCodec {
      * @param context {CodecRequestContext}
      * @param [documentHeaders] {Record<string, string>}   extra headers for
      *   document writes (the `Key-Epoch` stamp the codec seam applies)
+     * @param [baseline] {StoredDocument}   the stored document the transport's
+     *   first update replaces, which it pins that update to
      * @returns {WasTransport}
      */
-    #transportFor(context, documentHeaders) {
+    #transportFor(context, documentHeaders, baseline) {
         if (this.#transportFactory === undefined) {
             throw new NotSupportedError(`Collection "${this.#collectionId}" has no server behind it: this ` +
                 'codec was built for a local replica and holds no route to address ' +
@@ -591,14 +631,19 @@ export class EdvCodec {
                 'its `spaceId`, or by passing a `chunkSource` to the DocCipher ' +
                 '`decrypt` in place of a request context.');
         }
-        return this.#transportFactory({ context, documentHeaders });
+        return this.#transportFactory({
+            context,
+            documentHeaders,
+            ...(baseline !== undefined && { baseline })
+        });
     }
     /**
      * The plan for a binary payload over the single-document threshold, or for
      * a stream of any size: one EDV document plus its chunk resources, written
-     * by `EdvClientCore.insert({ doc, stream, transport })` over a transport
-     * built from the write context. The document id is minted before the plan
-     * is returned, so the caller can report it without waiting for the write.
+     * by `EdvClientCore.insert({ doc, stream, transport })` (or, over a stored
+     * document, `update`) on a transport built from the write context. The
+     * document id is minted before the plan is returned, so the caller can
+     * report it without waiting for the write.
      *
      * The `was` binding, the recipients and the write epoch are exactly the
      * single-document path's, and `additionalProtectedParams` carries the binding
@@ -609,69 +654,240 @@ export class EdvCodec {
      * payload, so that discriminator is what the read side routes on: a server
      * cannot mint it, and cannot suppress it to hide the chunks either.
      *
-     * The write is a create-if-absent: the first document write carries
-     * `If-None-Match: *`, so the plan also serves a write by id (`put()`) at an
-     * id that holds no document yet. A 412 there surfaces as the
-     * `PreconditionFailedError` the request path raised.
+     * With no prior document the write is a create-if-absent: the first
+     * document write carries `If-None-Match: *`, so the plan serves `add()` and
+     * a write by id (`put()`) at an id that holds no document alike. A 412 there
+     * surfaces as the `PreconditionFailedError` the request path raised. Over a
+     * prior document the first write is pinned to the pre-read's `ETag`, so a
+     * 412 is the ordinary lost-update signal, and the new chunks overwrite the
+     * prior ones index by index. Once the write is stored, a prior document
+     * that was itself chunked has every chunk beyond the new count reaped
+     * ({@link #reapChunks}): the new document's sealed count already makes them
+     * unreachable, and the reaping frees them.
      *
-     * The write is two-phase (`EdvClientCore.insert` writes the document, then
-     * streams the chunks), so a failure partway leaves a pending stub: a
-     * document whose sealed stream state is still `{ pending: true }`,
-     * undecryptable and listed. The plan therefore compensates: if the document
-     * was written and the write then failed, it best-effort deletes the stub
-     * before rethrowing. A stub that survives (the process died, or the cleanup
-     * failed) is recognized with {@link isPendingStub}. The EDV core cancels the
-     * caller's source stream on such a failure, so the source behind it is
-     * released.
+     * The write is two-phase (the EDV core writes the document, then streams
+     * the chunks, then writes the document again), so a failure partway leaves
+     * a pending stub: a document whose sealed stream state is still
+     * `{ pending: true }`, undecryptable and listed. A create compensates: if
+     * the document was written and the write then failed, the plan best-effort
+     * deletes the stub before rethrowing. A replace cannot restore the document
+     * it overwrote, so the stub is left in place for the next write at the id
+     * to replace, or for a `delete()`. A stub that survives is recognized with
+     * {@link isPendingStub}. The EDV core cancels the caller's source stream on
+     * such a failure, so the source behind it is released.
      *
      * @param options {object}
      * @param options.id {string}   the document id (minted, or the caller's)
      * @param options.stream {ReadableStream<Uint8Array>}   the payload, as the
      *   stream the EDV core re-chunks (never buffered whole by this codec)
      * @param options.meta {Record<string, unknown>}     the document meta to seal
+     * @param options.prior {PriorDocument | null}   the stored document this
+     *   write replaces, or `null` for a create
      * @returns {ChunkedWrite}
      */
-    #chunkedWrite({ id, stream, meta }) {
+    #chunkedWrite({ id, stream, meta, prior }) {
+        return this.#writePlan({
+            id,
+            meta,
+            prior,
+            write: async (transport) => {
+                const write = {
+                    stream,
+                    ...(this.#chunkSize !== undefined && {
+                        chunkSize: this.#chunkSize
+                    }),
+                    ...this.#sealParams({
+                        resource: id,
+                        hmac: this.#writeBlindingKey()
+                    }),
+                    transport
+                };
+                const stored = prior === null
+                    ? await this.#edv.insert({
+                        doc: { id, content: {}, meta },
+                        ...write
+                    })
+                    : await this.#edv.update({
+                        doc: {
+                            id,
+                            content: {},
+                            meta,
+                            sequence: prior.sequence,
+                            ...(prior.indexed !== undefined && {
+                                indexed: prior.indexed
+                            })
+                        },
+                        ...write
+                    });
+                // The EDV core reports the chunk count it sealed into the final
+                // document write; every stored chunk at or past it is the prior
+                // document's.
+                return stored.stream.chunks;
+            }
+        });
+    }
+    /**
+     * The plan for an inline document that replaces a chunked one: the sealed
+     * envelope is written as the single-document path would write it, pinned to
+     * the pre-read's `ETag` and stamped with the write epoch, and the prior
+     * document's chunks are then reaped ({@link #reapChunks}). One request
+     * cannot do both, which is why the write is a plan rather than an
+     * `EncodedWrite`. The envelope's sealed `meta.encoding` is inline, so a
+     * reader of the new document never looks for chunks, and the reaping only
+     * frees what is already unreachable.
+     *
+     * @param options {object}
+     * @param options.id {string}   the document id
+     * @param options.encrypted {IEncryptedDocument}   the sealed envelope
+     * @param options.meta {Record<string, unknown>}   the document meta sealed
+     *   in it, for the plaintext content type
+     * @param options.prior {PriorDocument}   the chunked document it replaces
+     * @returns {ChunkedWrite}
+     */
+    #replacingWrite({ id, encrypted, meta, prior }) {
+        return this.#writePlan({
+            id,
+            meta,
+            prior,
+            write: async (transport) => {
+                await transport.update({ encrypted });
+                return 0;
+            }
+        });
+    }
+    /**
+     * The lifecycle shared by every plan ({@link #chunkedWrite},
+     * {@link #replacingWrite}): a transport is built from the write context,
+     * stamped with the write epoch and, over a prior document, given the
+     * pre-read as its baseline; `write` runs on it and reports the chunk count
+     * the stored document covers; a failure is compensated and rethrown by
+     * {@link #chunkedWriteFailed}; a prior document that was chunked has the
+     * chunks past that count reaped ({@link #reapChunks}); and the result
+     * carries the validator the server acked the last document write with. The
+     * EDV core owns a chunked write and swallows the responses, which is why the
+     * transport is what reports the document write it made.
+     *
+     * @param options {object}
+     * @param options.id {string}   the document id
+     * @param options.meta {Record<string, unknown>}   the document meta sealed
+     *   in the write, for the plaintext content type
+     * @param options.prior {PriorDocument | null}   the stored document the
+     *   write replaces, or `null` for a create
+     * @param options.write {(transport: WasTransport) => Promise<number>}   the
+     *   write itself, resolving the chunk count the stored document covers
+     * @returns {ChunkedWrite}
+     */
+    #writePlan({ id, meta, prior, write }) {
         return {
             chunked: true,
             id,
             resourceContentType: meta.contentType,
-            // What the scheme-agnostic write path appends when it refuses this plan
-            // for a write by id over an existing document: only this codec knows
-            // why the payload needs several requests, and which low-level API
-            // writes one directly.
-            guidance: 'This payload is too large for a single encrypted document, or is a ' +
-                'stream, so it is stored as a document plus chunk resources. Drive ' +
-                'the write yourself with `EdvClientCore.update({ doc, stream, ' +
-                'transport })` over a `WasTransport`.',
             execute: async (context) => {
-                // The EDV core owns the write and swallows the responses, so the
-                // transport reports the document write it made: whether one landed at
-                // all (the cleanup decision below) and the validator the server acked
-                // it with.
-                const transport = this.#transportFor(context, writeHeaders({ epoch: this.#writeEpoch }));
+                const transport = this.#transportFor(context, writeHeaders({ epoch: this.#writeEpoch }), prior === null ? undefined : { id, etag: prior.etag });
+                let keep;
                 try {
-                    await this.#edv.insert({
-                        doc: { id, content: {}, meta },
-                        stream,
-                        ...(this.#chunkSize !== undefined && {
-                            chunkSize: this.#chunkSize
-                        }),
-                        ...this.#sealParams({
-                            resource: id,
-                            hmac: this.#writeBlindingKey()
-                        }),
-                        transport
-                    });
+                    keep = await write(transport);
                 }
                 catch (err) {
-                    throw await this.#chunkedWriteFailed({ err, id, transport });
+                    throw await this.#chunkedWriteFailed({
+                        err,
+                        id,
+                        transport,
+                        replace: prior !== null
+                    });
+                }
+                if (prior?.chunked) {
+                    await this.#reapChunks({ transport, id, keep });
                 }
                 const etag = transport.lastDocumentWrite?.etag;
                 return { id, ...(etag !== undefined && { etag }) };
             }
         };
     }
+    /**
+     * Reaps the chunks a replaced document left behind: lists the chunks stored
+     * under the document and deletes every one at or past `keep`, the chunk
+     * count the new document covers (`0` for an inline document). The listing,
+     * not the prior envelope's sealed count, says what is stored, so a pending
+     * stub's chunks (whose count was never sealed) and a longer stale set are
+     * reaped the same way. Each delete is pinned to the chunk `etag` the listing
+     * reported, where it reported one, so a chunk a later writer has since
+     * stored at that index is left to it; such a 412, and a chunk already gone
+     * (404), are not failures. The deletes run {@link REAP_CONCURRENCY} at a
+     * time. A listing that finds no resource (404) means another client deleted
+     * it since the write landed, and its chunks with it: nothing is left to
+     * reap.
+     *
+     * Runs only after the new document is stored, and the new document's sealed
+     * state already makes these chunks unreachable, so a failure here loses no
+     * data: it is reported as {@link EncryptionError} naming what is still
+     * stored. A later write at the id reaps again only if the stored document
+     * is still chunked (a pending stub, say); under an inline document the
+     * stale chunks stay until the resource is deleted, which takes them.
+     *
+     * @param options {object}
+     * @param options.transport {WasTransport}   the transport the write ran on
+     * @param options.id {string}   the document id
+     * @param options.keep {number}   the chunk count the stored document covers
+     * @returns {Promise<void>}
+     */
+    async #reapChunks({ transport, id, keep }) {
+        let stale;
+        try {
+            stale = (await transport.listChunks({ id })).filter(chunk => chunk.index >= keep);
+        }
+        catch (err) {
+            if (err instanceof NotFoundError) {
+                return;
+            }
+            throw this.#reapFailed(id, err);
+        }
+        const reapOne = async (chunk) => {
+            try {
+                await transport.deleteChunk({
+                    id,
+                    chunkIndex: chunk.index,
+                    ...(chunk.etag !== undefined && { ifMatch: chunk.etag })
+                });
+            }
+            catch (err) {
+                if (err instanceof NotFoundError ||
+                    err instanceof PreconditionFailedError) {
+                    return;
+                }
+                throw err;
+            }
+        };
+        // A fixed number of workers each take the next stale chunk until none is
+        // left, so at most `REAP_CONCURRENCY` deletes are in flight.
+        let next = 0;
+        const worker = async () => {
+            while (next < stale.length) {
+                await reapOne(stale[next++]);
+            }
+        };
+        try {
+            await Promise.all(Array.from({ length: Math.min(REAP_CONCURRENCY, stale.length) }, worker));
+        }
+        catch (err) {
+            throw this.#reapFailed(id, err);
+        }
+    }
+    /**
+     * The error a failed {@link #reapChunks} throws: the write landed, and what
+     * it did not reap is still stored.
+     *
+     * @param id {string}   the document id
+     * @param cause {unknown}   the failure
+     * @returns {EncryptionError}
+     */
+    #reapFailed(id, cause) {
+        return new EncryptionError(`The encrypted write of resource "${id}" is stored, but the chunk ` +
+            'resources of the document it replaced could not all be deleted. ' +
+            'They are unreachable through a read, which follows the chunk ' +
+            'count sealed in the stored document, but stay stored and counted ' +
+            'against quota; deleting the resource removes them.', { cause });
+    }
     /**
      * Compensates a failed chunked write and builds the error to rethrow. The
      * pending stub is deleted only when the transport reports it actually wrote
@@ -683,6 +899,11 @@ export class EdvCodec {
      * so its outcome only shapes the message. Deleting the document deletes the
      * chunks already stored under it.
      *
+     * A replace (`replace`) deletes nothing: the stub stands where the document
+     * it overwrote stood, and deleting it would turn a torn update into a
+     * missing resource. The message says the stub is there to be written over
+     * or deleted.
+     *
      * Three failures keep their own type, so a caller can dispatch on them as it
      * does for a single-request write. A 412 on the first document write (the
      * id is taken) or on the final document update (another writer replaced the
@@ -697,10 +918,12 @@ export class EdvCodec {
      * @param options.err {unknown}   the failure from the chunked write
      * @param options.id {string}     the document id the write targeted
      * @param options.transport {WasTransport}   the transport the write ran on
+     * @param options.replace {boolean}   whether the write replaced a stored
+     *   document (no stub delete) rather than created one
      * @returns {Promise<Error>}   the error to throw, carrying `err` as its cause
      *   when it wraps it
      */
-    async #chunkedWriteFailed({ err, id, transport }) {
+    async #chunkedWriteFailed({ err, id, transport, replace }) {
         // The transport rethrows a 412 or 409 as the named error the EDV core
         // dispatches on, with the typed HTTP error as its cause.
         const httpError = err instanceof Error &&
@@ -721,17 +944,26 @@ export class EdvCodec {
         }
         let outcome = 'The incomplete document and any chunks stored under it were deleted.';
         let removed = true;
-        try {
-            await transport.deleteDocument({ id, ifMatch: written.etag });
-        }
-        catch (deleteErr) {
+        if (replace) {
             removed = false;
             outcome =
-                deleteErr instanceof PreconditionFailedError
-                    ? 'Another writer has since replaced the document, so it was left ' +
-                        'in place.'
-                    : 'The incomplete document could NOT be deleted and is still ' +
-                        'stored; delete it before writing again.';
+                'The document it replaced is gone, and the resource now holds this ' +
+                    "write's pending stub, which no read can open: write the resource " +
+                    'again, or delete it.';
+        }
+        else {
+            try {
+                await transport.deleteDocument({ id, ifMatch: written.etag });
+            }
+            catch (deleteErr) {
+                removed = false;
+                outcome =
+                    deleteErr instanceof PreconditionFailedError
+                        ? 'Another writer has since replaced the document, so it was ' +
+                            'left in place.'
+                        : 'The incomplete document could NOT be deleted and is still ' +
+                            'stored; delete it before writing again.';
+            }
         }
         if (httpError instanceof QuotaExceededError ||
             httpError instanceof ConflictError) {