SSE delivery lab

Each link is a complete configuration. Open in a real, visible window — Safari marks automation windows document.hidden and throttles them, and every report states whether the page stayed visible for the whole run. Start from one, change any control, and the address bar becomes that configuration — so a tweaked scenario is still a link you can paste into a bug.

Transport

transportHow the page reads the stream. THE decisive lever. Reading the WebKit source, fetch goes through FetchBodySource, which appears to stash bytes without draining them on read, while EventSource parses directly with no queue. Measured: EventSource was unaffected in every run.

modeWhat drives emission. timer: the server emits on its own clock. command: one POST per event, matching how an app pushes on interaction. manual: nothing fires until you click.

decoupleDrain the stream into a queue and apply on a separate task. Kept as a NEGATIVE result: a separate task is still the same thread, so read() still cannot resolve. It does not help.

pokeAfter applying, the page tells the server it is idle and the server writes a few bytes. Best latency in isolation, because the client knows when it is free while a server timer must guess.

Encoding & shape

encContent-Encoding for this stream only. Measured NOT to matter: identity strands as badly as br at the same payload size. Encoding only modulates how many delivery callbacks a burst produces. zstd is here because WebKit bug 311087 reports a separate, encoding-specific stall for it — this is the knob to test whether that is the same defect or a different one. It works in Safari even though Safari never advertises zstd: the lab forces the encoding server-side, and WebKit decodes it anyway (verified).

ctypeOverrides the response Content-Type. Kept as a NEGATIVE result: neither application/json nor octet-stream avoids the stranding, though CFNetwork is known to branch on this header elsewhere.

shapejson puts each event on a single data line. datastar uses the real patch framing — a selector and mode preamble, then one elements line per line of HTML.

linesHow many data lines to split the payload across. A real morph sits near 138. Measured irrelevant: line count does not change the outcome.

applyWhat the reader does with each payload. morph runs Idiomorph, the algorithm Datastar patches with. Anything that occupies the main thread will do.

Timing

slowMilliseconds the reader blocks the main thread after each record, standing in for the cost of applying an update. A dose-response with no safe floor: even 0 strands about 8%.

spreadMilliseconds between the events of a burst. Spacing is not the same as merging — below about 500ms it does not help at all.

burstEvents written back-to-back per trigger. THE decisive variable: 1 is clean at any interval, 2 strands. Holding busy time constant and halving this reverses the outcome.

gapMilliseconds between triggers. A stranded event is released by the next write, so this effectively sets the observed delay — 800ms gap, 800ms delay; 5s gap, 5s delay.

Payload

sizePayload bytes per event. Shifts the distribution but has no clean threshold — a size sweep is non-monotonic, with 32KB stranding where 64KB does not.

countNumber of triggers. Total events emitted is count x burst.

noiseInterval in ms for background POSTs contending for the same HTTP/2 connection, as an app's throttled scroll traffic would. 0 disables.

Output

writemany gives one encoder flush per event; one puts the whole burst in a single flush. Batching does NOT help — the record count is what matters, not the flush boundary.

padkindrandom padding is incompressible; repeat compresses to almost nothing. Only incompressible padding has any effect, which places the threshold on compressed input bytes.

padBytes of trailing SSE comment after each burst. 4KB does nothing; 64KB drops the p90 from 4.9s to 68ms but never fully clears it.

flushMilliseconds before a small trailing comment is sent. 150 fires while the consumer is still busy and achieves nothing; 250 partly helps; 500 works completely.

flushbytesSize of that trailing comment. 16 bytes is enough — what releases a stranded fragment is input arriving LATE, not input arriving large.

idle