Hyperlang 0.2.0

Public API

• Breaking: Hyperlang.module(program) is removed. It returned a code-slot
  module name that was not pinned until an actor's init/1 ran. A release
  or reclamation in between made the start fail with undef, or made the
  actor run another program that had reused the slot.
• Hyperlang.Actor is the public GenServer callback module for program
  actors. Its init argument is {spec, program, arg}, and child_spec/1
  takes the same tuple (id Hyperlang.Actor, restart :permanent; give
  each of several actors under one supervisor its own id with
  Supervisor.child_spec/2). The actor's init/1 loads the program and
  pins its slot in one step, so the actor always runs exactly that program.
  Bad init arguments stop it with {:hyperlang, %{category: :api}}.
• Migration:
  GenServer.start(Hyperlang.module(program), {spec, arg}) becomes
  GenServer.start(Hyperlang.Actor, {spec, program, arg}), and
  GenServer.start_link likewise.
• Hyperlang.load/1 loads a program ahead of its first use and returns
  :ok or {:error, sanitized_error}. Migration: a preloading
  Hyperlang.module(program) call becomes :ok = Hyperlang.load(program).

Behaviour

• Generated modules no longer export OTP callbacks; GenServer.start/2 on
  one fails with undef. An actor's process module is Hyperlang.Actor.
• An actor calls handle_info/2 and terminate/2 only when the program
  exports them. A message to a program without handle_info/2 is dropped
  without a log entry. A missing handle_cast/2, handle_call/3 or
  handle_continue/2 stops the actor with the typed run/4 error
  (category: :export) instead of undef.
• An actor's init/1 no longer stops with code_pool not_loaded. It stops with
  code_pool when the program's window has no free slot, or with
  pin_limit at the scope bound (checked before any load).
• The compile for a load done by init/1 is charged to the actor's scope.

Isolation verification

• Heap pressure has a tested bound: with 16 heap-pressure tenants in
  :background scopes with a churn quota, the latency-class victim's p99
  stays within 3x its unloaded p99. The R2-4 paired measurement now spreads
  loaded blocks over a whole quota refill period, so rate-limited bursts are
  measured.
• Corrected noisy-neighbour figures. The sequential matrix and the comparison
  bench sampled the victim only before the first quota refill, so they never
  saw the noisy tenants' refill bursts. Requests are now spread evenly over
  four refill periods. With 16 tenants in background scopes the victim's p99
  slowdown is 1.7x (CPU), 2.1x (heap) and 1.9x (compile), and the comparison
  figure is 1.65x; 0.1.0 published 1.1x, 1.1x, 1.0x and 1.37x.

Package

• The Hex package links to this changelog.