chore(release): v0.2.0

[kalaspuff]

Lean, fast defaults (breaking)

Without any tuning, clients and consumers now perform like a hand-tuned 0.1 configuration
(ADR 0008):

• Messages are published with the raw envelope: the body is the encoded payload (e.g. the
  JSON document); fanferro.id and fanferro.type (typed messages) travel as message attributes,
  content-type only for payloads that are not JSON. CloudEvents is opt-in:
  Client(envelope=fanferro.envelopes.CloudEvents()). Consumers still read CloudEvents (JSON and
  protobuf), so messages from 0.1 producers keep working.
• MD5 verification of sent and received messages is opt-in: Client(verify_md5=True).
• consumer.stop() returns at once: ShutdownPolicy.abort_receives defaults to True (running
  handlers still drain). SQS may still hand the aborted long poll up to batch_size messages per
  poller; they are redelivered after their visibility timeout.
• Full receives: a poller waits until min_receive_batch handler slots are free (default:
  batch_size; FIFO queues: 1), for at most receive_batch_wait (0.1 s) after the first one, and
  prefetch defaults to one receive batch per poller (min(batch_size, concurrency) × pollers).
• Slots are freed when settlement starts: a handler that returns hands its slot to the next
  message while the delete is in flight; settlements in progress are bounded separately.
• Batched acks: deletes and visibility changes wait up to settle_linger (5 ms) to fill a
  batch of ten, and are sent at once when every running handler awaits an explicit settlement
  in that batch.
• Automatic acks are not awaited per message: the engine deletes the message in the
  background (consumer.stop() waits for it). consumer.stats()["processed"] and the
  message.processed event count handlers that succeeded; stats()["engine"]["messages_acked"]
  counts confirmed deletes, and a failed delete is reported as a message.ack_failed event. Explicit
  settlements (await message.ack()...) still resolve once SQS confirmed them.
• consumer.stop() and client.close() no longer wait up to half a second for the dispatcher or
  the event pump, and events emitted just before close() are no longer lost.
• Per-message work in the consumer and publisher is lighter: settlement awaitables are resolved
  in batches, event checks, type names and encoders are cached, and messages are no longer copied
  while being decoded.

Migrating from 0.1:

• Upgrade consumers before producers. 0.1 consumers read 0.2 messages (typed routes, message ids
  and filter policies on fanferro.type work), but they do not imply JSON for raw messages:
  untyped routes see JSON scalars as bytes, and CloudEvents metadata (message.source,
  subject, headers) is absent.

• Services that parse fanferro messages as CloudEvents themselves keep producing them with
  Client(envelope=fanferro.envelopes.CloudEvents()).

• The 0.1 behaviour is one configuration away:

  client = fanferro.Client(envelope=fanferro.envelopes.CloudEvents(), verify_md5=True)
  consumer = client.consumer(
      queue,
      min_receive_batch=1,
      prefetch=0,
      settle_linger=0,
      shutdown=fanferro.ShutdownPolicy(abort_receives=False),
  )

Features

• consumer: Full receive batches (min_receive_batch, receive_batch_wait, default prefetch) (feat(consumer): full receive batches (min_receive_batch, receive_batch_wait, default prefetch) #5) (feat(consumer): full receive batches (min_receive_batch, receive_batch_wait, default prefetch) #5)
• [breaking] Lean defaults - raw envelope, opt-in CloudEvents and MD5 checks, immediate stop (feat!: lean defaults - raw envelope, opt-in CloudEvents and MD5 checks, immediate stop #6) (feat!: lean defaults - raw envelope, opt-in CloudEvents and MD5 checks, immediate stop #6)

Bug fixes

• consumer: Wake the dispatcher and the event pump immediately when stopping (fix(consumer): wake the dispatcher and the event pump immediately when stopping #2) (fix(consumer): wake the dispatcher and the event pump immediately when stopping #2)

Performance

• consumer: Free handler slots when settlement starts and bound the settling backlog (perf(consumer): free handler slots when settlement starts and bound the settling backlog #3) (perf(consumer): free handler slots when settlement starts and bound the settling backlog #3)
• consumer: Batch acks and retries (full batches at once, 5 ms linger) (perf(consumer): batch acks and retries (full batches at once, 5 ms linger) #4) (perf(consumer): batch acks and retries (full batches at once, 5 ms linger) #4)
• Cheaper per-message consume and publish paths (perf: cheaper per-message consume and publish paths #7) (perf: cheaper per-message consume and publish paths #7)