UMP and decoding
Part of SABR in the extractor. YouTube wraps SABR responses in UMP (Ultra-Minimal Playback). UMP frames are not protobuf; the payload of most control parts is protobuf and is summarized or decoded by SabrResponseDecoder.
UMP framing
UmpReader reads a sequence of parts:
[type: UMP-varint][size: UMP-varint][payload: size bytes] ...readPayloadsUntil(InputStream, consumer) reads one part at a time from the network. readAll(byte[]) parses an already-buffered body, which is useful for small responses and tests. A clean EOF at a part boundary ends the stream; truncated headers or payloads raise SabrProtocolException.
UMP varints choose their length from the high bits of the first byte:
| First byte | Total bytes | Value |
|---|---|---|
0x00–0x7F | 1 | byte value |
0x80–0xBF | 2 | (b0 & 0x3f) + 64·b1 |
0xC0–0xDF | 3 | (b0 & 0x1f) + 32·(b1 + 256·b2) |
0xE0–0xEF | 4 | (b0 & 0x0f) + 16·(b1 + 256·(b2 + 256·b3)) |
0xF0–0xFF | 5 | next four bytes, little-endian |
UmpReader.UmpPart exposes the type, size and payload. Public data access is defensive; the streaming decoder uses the raw array internally to avoid copying large media parts.
Part identifiers
SabrResponseDecoder recognizes these identifiers. Rich control messages are kept as summaries or raw bytes in YoutubeSabrResponse; they are not modeled as one Java class per part.
| ID | Constant | Current handling |
|---|---|---|
| 10–12 | ONESIE_* | generic summary |
| 20 | MEDIA_HEADER | decode SabrMediaHeader |
| 21 | MEDIA | count bytes by header id |
| 22 | MEDIA_END | close a header id |
| 30–34 | config/live hints | generic summary |
| 35 | NEXT_REQUEST_POLICY | keep raw policy and read backoff field 4 |
| 36–38 | ustreamer metadata | generic summary |
| 42 | FORMAT_INITIALIZATION_METADATA | decode generated metadata |
| 43 | SABR_REDIRECT | keep the redirect URL |
| 44 | SABR_ERROR | decode type/code summary |
| 45 | SABR_SEEK | generic summary |
| 46 | RELOAD_PLAYER_RESPONSE | set reloadRequested |
| 47–51 | playback/format controls | generic summary |
| 52–56 | request controls | generic summary |
| 57 | SABR_CONTEXT_UPDATE | keep raw context update |
| 58 | STREAM_PROTECTION_STATUS | decode status and max retries |
| 59–65 | context/cache/connection controls | keep summary; 65 is prewarm |
| 66–67 | debug/snackbar | generic summary |
Unknown identifiers are recorded in the bounded diagnostic summary. Malformed control parts are retained as bounded diagnostics so valid media in the same response is not discarded.
The decode paths
SabrResponseDecoder.decode(byte[]) parses all parts first and then dispatches them. The normal application path uses SabrStreamingResponseReader.read(InputStream, consumer, startConsumer, spool): it keeps control parts in memory and feeds MEDIA_HEADER, MEDIA, and MEDIA_END directly to SabrMediaSegmentCollector.Incremental. Completed segments are delivered as soon as their MEDIA_END arrives, so a large 4K body does not have to remain in the heap.
Both paths count media bytes by header id. YoutubeSabrResponse.getIntegrityIssues() then detects duplicate headers, missing media, length mismatches, missing ends, media without a header, and ends without a header. The session retries only the classified incomplete-media cases within its bounded limit.
The decoded response
YoutubeSabrResponse exposes the HTTP response code and content type, UMP parts, completed segments, initialization metadata, live metadata, context updates, redirect/error/reload flags, protection status, backoff, byte counters and bounded summaries. summarizeForDiagnostics() reports structure and sizes, not PO-token or cookie contents.