Repository navigation
Robotweax SRT 0.3.4 adds an epoll wake call, hardens connection setup, group membership and file receive, reduces wake-ups on sparse traffic, and adds three default-off experimental profiles: MPEG-TS null packet deletion, pinned-key peer authentication and Linux UDP batching.
Release status
Final-tag CI, package qualification, the protected independent rebuild and signing are complete. The OpenSSL and BCrypt Windows SDK installers are Authenticode-signed by Robotweax GmbH. Their downloaded bytes match the signed CI artifact and the attached SHA256SUMS manifest; tag provenance has been verified.
Changes
robotweax_srt_epoll_wake(eid)lets another thread end onesrt_epoll_waitorsrt_epoll_uwaitwithout registering a system socket or closing an SRT socket. A wake is not distinguishable from a timeout by return value. Programs that never call it see no change.- A connection group retains at most
ROBOTWEAX_SRT_MAX_GROUP_MEMBERSmembers, pending and connected together. The limit is a build-time setting from 1 to 65536 with a default of 64; deployments that relied on larger groups must raise it when building. - Setup input is validated before cryptographic or metadata state changes. Once an attempt is bound, a packet whose peer cookie, socket ID or flow window differs is ignored and the original attempt continues within its deadline, instead of failing the attempt.
- File receive keeps one file handle across network waits and delayed truncation, so a rename after opening cannot redirect truncation or writes. A final symbolic link or Windows reparse point is rejected; initial open failures return
SRT_EWRPERMimmediately. - A channel with buffered DATA that has received no datagram for 20 ms now waits for its earliest protocol deadline, capped at 10 ms, instead of polling every 2 ms. This changes wake-up timing for sparse traffic on every build. No general throughput or CPU improvement is claimed.
- Per-packet paths drop redundant zero-fills and a per-byte buffer-overlap test.
ROBOTWEAX_SRT_ENABLE_LTOis a new opt-in build option. - In the default-off hybrid control profile, freshness refreshes are scheduled from the confirmation origin, so delayed but valid replies no longer let a confirmed policy expire between refreshes.
Experimental profiles
These additions are default-off, use provisional identifiers and are outside the supported security and interoperability boundary.
- Null packet deletion.
SRTO_ROBOTWEAX_NPD_POLICYremoves canonical MPEG-TS null packets at the sender and restores them bit-exactly at the receiver on individual Live Caller/Listener sockets. Stock peers do not implement it. - Pinned-key peer authentication. Builds with
ROBOTWEAX_SRT_ENABLE_EXPERIMENTAL_PEERAUTHadd caller/listener authentication with pinned P-256 identities, for OpenSSL only. The cryptographic design has not been independently reviewed and no support or security claim is made. A listener completes one setup at a time; this is a documented limit, tracked in #406. - Linux UDP batching. Builds with
ROBOTWEAX_SRT_ENABLE_EXPERIMENTAL_UDP_BATCHINGuserecvmmsgandsendmmsg;ROBOTWEAX_SRT_ENABLE_EXPERIMENTAL_UDP_SEGMENTATIONaddsUDP_SEGMENT. Selection is compile-time only.
Compatibility and scope
Project version 0.3.4, shared-library ABI 0.3, compatible SRT API 1.5.7. All 85 C exports and the existing public C layouts of 0.3.3 remain, and robotweax_srt_epoll_wake is added, for 86 in total. New socket options and the new call require matching headers and library. Native source-tree C++ consumers must rebuild against matching sources; ABI 0.2 applications also require rebuilding.
Long-horizon compatible-mode KMREQ replay limits, shared multicast-key trust and revocation limits, and the experimental ACP1 and hybrid restrictions are unchanged from 0.3.3. A lost final Live packet is still not guaranteed to recover without later DATA. Loopback and virtual-host measurements do not establish physical-network, WAN or general throughput guarantees.
See the complete release notes, null packet deletion, peer authentication draft, UDP I/O backend draft, 0.3.4 security delta and release contract.
Source and qualification
- Release commit:
87742777290e6a77e38672ca35701f7184b75526. - Immutable product archive:
6009c6430b46e8691181bbd8690bf36c3ef0d661, the first and only frozen product of this release. - Archive SHA-256:
66ab41f65523fce7cea66c5ac54b44f3f8d7b5a87586755b00ee91a70d870116. - Pre-tag full CI passed on the first attempt on the recipe commit (44692277103), including the media integrations and all OBS platform jobs. On the merged head it passed on the second attempt (44735673903); see the first failures below.
- Exact-archive Linux packages and package managers, and the exact-product Lab smoke, interop and telemetry requests passed. A commit-bound local macOS hybrid soak passed.
- The two residual risks accepted for 0.3.3 carry over unchanged and are not marked fixed: the historical macOS identity-rendezvous timeout and the intermittent OBS desktop reconnect failure.
- Three intermittent hosted test failures have no established cause:
- macOS: the ARQ-first recovery sample delivered 159 of 160 messages, in a pull-request run and in the first attempt on the merged head. A hosted comparison of 1,200 iterations per source failed once on the product and three times on
v0.3.3, so the failure is not introduced by 0.3.4. - Windows: a Python-only test of the example process helper timed out in the first attempt of two pull-request runs.
- Windows:
srt-live-transmit -versiondid not return within 60 seconds in the first attempt of the final-tag CI run.
- macOS: the ARQ-first recovery sample delivered 159 of 160 messages, in a pull-request run and in the first attempt on the merged head. A hosted comparison of 1,200 iterations per source failed once on the product and three times on
- Three corrections from the security delta have no test that observes them: expected-tag erasure, batch receive hardening and the group enrolment rollback.
- The record is in
packaging/qualification-0.3.4.mdandpackaging/acceptance-0.3.4.json. The tagged record predates the merge; the merged-head and final-tag results, including the second Windows helper-test timeout, the merged-head macOS failure and thesrt-live-transmithang, are stated here and not in that file.
Final-tag release evidence
- Full source CI and Required CI gate passed on the release commit on the second attempt. In the first attempt only "srt-live-transmit / windows-latest" failed; that job and the gate were re-run.
- Package-manager qualification and Linux package qualification passed.
- Windows SDK qualification, protected rebuild, signing and release draft passed in one workflow attempt. Exactly one protected independent rebuild ran for this release; the optional unsigned rebuild diagnostic was skipped.
- Both downloaded release installers match signed CI artifact
14640243579byte-for-byte. GitHub asset digests and both installer checksums inSHA256SUMSalso match. - The unchanged annotated tag object is
559c0c4a0ba0ddbb0d65a62eebbe127f5187f2a5;tag-provenance.jsonbinds it to the release commit. Use the attachedSHA256SUMSto verify installer downloads.