v0.11.7

Many important fixes for RTS/DB and the language in general!

Added

• Bash completion is now part of the Debian packages & brew formula

Changed

• actondb now uses a default value for gossip port of RPC port +1 [#913]
  • The gossip protocol only propagates the RPC port & parts of the
    implementation has a hard-coded assumption that the gossip port has a +1
    offset
  • In order to avoid configuration errors, the default gossip port is now RPC
    port + 1 and if another gossip port is explicitly configured, an error log
    message is emitted on startup.
  • While this is marked as a change, it could really be considered a fix as any
    other configuration of the system was invalid anyway.

Fixed

• Fixed include path for M1
  • /opt/homebrew/include added to header include path [#892]
  • Actually fixes builds on M1!
  • This has "worked" because the only M2 where Acton was tested also had header
    files in /usr/local/include but on a fresh install it errored out.
• Fix up-to-date check in compiler for imported modules from stdlib [#890]
• Fix seed arg parsing in actondb that lead to "Illegal instruction" error
• Fix nested dicts definitions [#869]
  • Now possible to directly define nested dicts
• Avoid inconsistent view between RTS & DB in certain situations [#788]
  • If an RTS node was stopped & quickly rejoins or if a transient partition
    happens and the gossip round does not complete before the partition heals.
  • We now wait for gossip round to complete.
  • This ensures that local actor placement doesn't fail during such events.
• Fix handling of missed timer events [#907]
  • Circumstances such as suspending the Acton RTS or resuming a system from the
    database could lead to negative timeout, i.e. sleep for less than 0 seconds.
  • The libuv timeout argument is an uint64 and feeding in a negative signed
    integer results in a value like 18446744073709550271, which roughly meant
    sleeping for 584 million years, i.e. effectively blocking the RTS timerQ.
  • It's now fixed by treating negative timeouts as 0, so we immediately wake up
    to handle the event, however late we might be.
• Timer events now wake up WT threads after system resumption [#907]
  • Worker Threads (WT) are created in NoExist state and should transition
    into Idle once initiated, however that was missing leading to a deadlock.
  • This was masked as in most cases, a WT and will transition into Working
    once they've carried out some work and then back into Idle
  • wake_wt function, which is called to wake up a WT after a timer event is
    triggered, wakes up threads that are currently in Idle state, if they are
    in NoExist, it will do nothing.
  • If there is no work, such as the case after system resumption from the DB,
    WTs will stay in the NoExist state and then wake_wt will do nothing, so
    the system is blocked.
  • WT now properly transition into Idle.
• Only communicate with live DB nodes from RTS DB client [#910] [#916]
  • When the RTS communicates with the DB nodes, we've broadcast messages to all
    servers we know about. If they are down, they've had their socket fd set to
    0 to signal that the server is down. However, fd=0 is not invalid, it is
    stdin, so we ended up sending data to stdin creating lots of garbage output
    on the terminal.
  • fd -1 is used to signal an invalid fd, which prevents similar mistakes.
  • The DB node status is inspected and messages are only sent to live servers.
• Avoid segfault on resuming TCP listener & TCP listener connection [#922]
  • Invalidate fds on actor resumption [#917]
• Remove remaining ending new lines from RTS log messages [#926]
• Remove ending new lines from DB log messages [#932]

Testing / CI

• Rewritten RTS / DB tests [#925] [#929]
  • More robust event handling, directly reacting when something happens, for
    example if a DB server segfaults or we see unexpected output we can abort
    the test
  • Now has much better combined output of DB & app output for simple
    correlation during failures
  • Test orchestrator now written in Acton (previously Python), at least async
    IO callback style is better supported to directly react to events...