timely-v0.30.0

This release adds opt-in spill-to-disk support for the zero-copy network allocator, makes Bytes Sync (a soundness fix), and continues trimming compile-time monomorphization sprawl.

Breaking changes

• Bytes is now Sync, and byte buffers must be Send. timely_bytes::arc::Bytes now implements Sync in addition to Send. To make both impls sound, BytesMut::from now requires its payload to be Send. This is a breaking change to timely_communication: BytesRefill::logic now produces Box<dyn DerefMut<Target=[u8]> + Send> rather than Box<dyn DerefMut<Target=[u8]>>. Custom refills whose buffer type wraps a raw pointer (e.g. NonNull) must assert unsafe impl Send on that type. (#800)
• ToStreamBuilder exposes the item type via the Item associated type instead of a trait-level generic, and the container builder moves to a method-level generic. This enables method-call syntax: (0..3).to_stream_with_builder::<_, CapacityContainerBuilder<_>>(scope) instead of the UFCS form ToStreamBuilder::<CapacityContainerBuilder<_>>::to_stream_with_builder(0..3, scope). (#792)

Added

• Spill-to-disk for the zero-copy allocator (timely_communication). A new allocator::zero_copy::spill module lets a MergeQueue shed resident bytes under memory pressure. It composes three pluggable traits: SpillPolicy (whether and how to reshape a queue on each extend), BytesSpill (where spilled bytes go — file, object store, mock, …), and BytesFetch (reads them back). The shipped threshold::Threshold policy spills the middle of a queue once resident bytes exceed configurable threshold/reserve/budget knobs. It is opt-in via the spill hook on the communication configuration and defaults to None, so existing deployments are unaffected. See the spill_stress and spill_compare examples. (#789, #791)
• logging::Registry::names iterates the names of the currently bound loggers. (#795)

Other

• Updated the columnar dependency to 0.13.
• Hoisted the per-update rebuild check out of MutableAntichain::update_iter into a requires_rebuild helper generic only over the timestamp, reducing monomorphization sprawl (~1600 fewer LLVM lines on the event_driven example). (#797)