Cancel jobs instead of dropping them on controller loss

A worker that stays without controller heartbeats longer than controller_wait_timeout now cancels its jobs with a typed AbandonedJob error and keeps their runtime records, instead of dropping them and restarting the same job ids in place (which silently lost accepted-but-unconfirmed input messages, leaving upstream distributors waiting forever). The controller learns about the cancelled jobs through regular status reports and recreates them with new ids without the failure backoff. The execution spec is no longer reset on disconnect, so the data plane keeps working while the controller is away.
 Cancel jobs instead of dropping them on controller loss
commit_hash:faf38ecdc0637d5ee86d50de3b5e22d6e148ae70

0 files changed, 0 insertions, 0 deletions