mirror of
https://github.com/block/buzz.git
synced 2026-10-31 22:41:12 +00:00
937a4726eb
reconcile_buzz_mesh_join dialed only the first non-visible target,
sorted by reporter pubkey. dial_endpoint_addr is a background bootstrap
hint (Ok does not mean the connection landed), so one permanently
unreachable member that sorts first starves every member after it: two
healthy machines each dialed the same dead third node every 15s and
never dialed each other.
Now every non-visible member gets dialed each poll; failures aggregate
so one bad endpoint no longer aborts the pass.
Verified: mesh_llm scoped tests 81/82 pass (1 ignored); full --lib run
2855 passed with one pre-existing failure
(inherited_shared_compute_translates_to_supported_agent_transport,
fails identically on clean HEAD e0933fd4b).
Co-authored-by: Thomas Petersen <thomasp@squareup.com>
Signed-off-by: Thomas Petersen <thomasp@squareup.com>