Crate ts_l3 

The L3 storage tier’s Host-only executor, owned by the engine host (docs/design/async-l3-cache.md).

The C++ scheduler plans each request’s storage operations and the control plane agrees on their outcomes across the replica (ts-control’s l3); this crate runs the operations on this rank: bounded queues, worker threads, the backend’s health, and the stores — the Mooncake store client, and an in-memory one for tests. Workers read and write the Host pool the data plane allocated, by address; no thread here takes the GIL, calls the scheduler or touches the GPU.

Re-exports

pub use executor::monotonic_clock;

pub use executor::Clock;

pub use executor::Executor;

pub use executor::ExecutorConfig;

pub use executor::ExecutorError;

pub use executor::Namespace;

pub use executor::Operation;

pub use executor::StorageCompletion;

pub use executor::StorageKey;

pub use executor::StoragePoll;

pub use executor::StoragePressure;

pub use layout::GroupLayout;

pub use layout::HostLayout;

pub use layout::LayoutError;

pub use layout::Range;

pub use settings::Settings;

pub use settings::StoreSettings;

pub use store::process_memory_store;

pub use store::HostBuffer;

pub use store::IoResult;

pub use store::MemoryStore;

pub use store::Store;

pub use store::StoreError;

Modules

executor

The executor: bounded queues of storage jobs and the worker threads that run them against the store (docs/design/async-l3-cache.md).

keys

Object keys: a namespace per deployment and weight generation, and one object per packed Host cache block.

layout

Where a scheduler Host block lives in the shared pinned Host allocation, and whether this rank owns it.

mooncake

The Mooncake store as an L3 store: zero-copy gets and puts against the registered Host pool, as SGLang’s HiCache and vLLM’s connector use it.

settings

Everything an executor starts from, as the data plane resolves it at startup (cache/l3/config.py’s L3Settings): the limits, the namespace, the Host pool and the store.

store

The byte stores packed Host cache blocks go to, and the in-memory one.