Skip to content

v2.11.1

Latest

Choose a tag to compare

@github-actions github-actions released this 16 Nov 14:51
· 86 commits to main since this release

Fixed

  • Under journald (systemd), every logged exception showed the log message where the exception's message belongs, so the actual error was missing from the journal, e.g. Read Loop error (IO::Error). A cause line also got a stray <3> priority prefix. Exceptions are now logged with their own message, as with stdout logging, with line breaks in it escaped so it stays one journal entry #2698
  • A delayed message that came due during a shutdown, after the shutdown closed its destination queue, was lost. The broker now keeps the message in the delayed exchange and releases it after the restart #2694
  • A message that was being written to a queue when the broker was killed could be delivered after the restart with part of its body missing (zero bytes), instead of being dropped like any other publish that was never confirmed. A message's timestamp is now written last, after its body, and a message without a timestamp is dropped at start. The segment format is unchanged #2735
  • In a priority queue, messages of different priorities could share a delivery count. With x-delivery-limit, messages were dead-lettered before or after their limit, or dropped early without a dead letter exchange, and a first delivery could be flagged as redelivered. With automatic retries, messages were dead-lettered before their configured number of retries and retried with the wrong backoff. Each message now has its own delivery count #2744
  • basic.recover with requeue=false closed the connection when the channel had an open consumer. Also, a basic.get message requeued by basic.recover still counted as a basic.get message, so the channel's consumers got one message more than the global prefetch for each such message. And basic.recover with requeue=true dropped messages already acked in an open transaction from the channel's unacked list, so tx.commit didn't ack them and they stayed unacknowledged on the queue until a restart #2717
  • PUT /api/queues/:vhost/:name/pause, /resume and /restart didn't check the user's permissions on the queue, so any management user with access to the vhost could pause, resume or restart any queue in it. They now require configure permission on the queue, like delete does #2690
  • On a queue with x-single-active-consumer, the next consumer could get no messages after the active consumer cancelled, while the queue had messages ready. The handoff was lost when it happened while the next consumer's deliver loop started. Now only the active consumer runs a deliver loop, and the handoff starts it #2693
  • After an exchange was deleted and an exchange with the same name declared again, a restart brought back the deleted exchange's bindings on the new exchange, so it routed messages to queues and exchanges nobody had bound it to #2756
  • A follower's full sync compared its files with the leader's in quadratic time, without letting other fibers run, which blocked its client proxies, HTTP API and replication for seconds with tens of thousands of files, and about a minute with 100,000. The comparison is now linear and lets other fibers run every 1,024 files #2755
  • A queue declared while the server was low on disk space was still created, even though the declare was refused with PRECONDITION_FAILED #2560
  • A lower priority consumer that was closed while it waited for a higher priority consumer to run out of capacity kept waiting. If the higher priority consumer stayed idle, the deliver loop fiber, the closed consumer and its channel stayed in memory until the queue closed. The wait now ends when the consumer is closed #2636
  • The Upstream column of the federation links table was always empty #2574
  • The exchange page's rates chart had room for a single y axis tick and cut off its legend labels at common window widths. Charts now have a minimum height, and the legend's value column only takes the room its values need #2572
  • The Bindings counter and footer on the overview sat lower than the others whenever the other cards' links wrapped, e.g. with a long uptime #2573
  • Since 2.11.0, a connection that the broker closed while it held a pooled socket read buffer, such as after a failed AMQP login, didn't return the buffer to the pool. The socket_buffer_pool_* metrics then counted it as still in use, and the pool allocated a new buffer instead of reusing one. The buffer is now returned when the connection closes #2692