Crate rushdb_storage 

Ordered KV substrate shared by every node: buffered transactions over a committed B-tree view, atomic redb publication of expanded op sets, a serialized command log, multi-version snapshot reads, and serializable-optimistic dependency validation.

Re-exports

pub use planner_stats::PlannerStats;

Modules

backend

Narrow storage-substrate seam (research D6, plan W3).

cache

Byte-accounted logical caches.

history_v2

Bounded, append-friendly MVCC history blocks.

log_store

Segmented, append-only write-ahead log (log-first).

owner_sets

Owner sets of declared-index leaves, cached per snapshot (READ-2).

planner_stats

Evictable planner-statistics cache (DC1 “on demand, evict when idle”).

read_accounting

Opt-in read accounting for profiling: counts the backend’s point reads, range opens and seeks per key family (keys::family_label). Off by default; when off, a read costs one relaxed atomic load. take returns every thread’s reads (profile one workload at a time); take_local returns only the calling thread’s, which stays exact while other work runs concurrently (tests).

resources

Shared managed-byte accounting. Reservations are ownership tokens, not allocations, an RSS limit, or a substitute for charging each consumer.

tenant_dir

Tenant directory: scope digest ↔ 4-byte tenant ordinal.

wal

Record framing for the log-first storage substrate. This is the write-ahead log: crate::log_store appends these framed records to segment files, and that durable append is the commit’s ack boundary (not the buffered redb materialization). See the manual’s docs/manual/architecture/storage-engine.md.

Structs

BackendLockTimes

Cumulative backend database-lock timings, summed over every shard.

DependencySet

Serializable-mode read dependencies recorded against a transaction. Zero-width entries are point reads; otherwise half-open [start,end).

OverlayView

A pinned base snapshot plus an uncommitted key overlay.

PreparedCommand

Prepared metadata for one admitted command. The publisher keeps this alongside expanded operations so an admission overlay can validate read/write conflicts without re-expanding the command.

ReadCancellation

Shared cancellation flag for one or more protected read views.

ReadLimits

Per-read byte ceilings enforced when copying values out of the store.

ReadPin

One commit index’s retained read lease.

ReadView

Owns all resources needed to read; it can outlive the Store and be shared with asynchronous providers without self-referential borrowed lifetimes.

ScanBatch

One batch of scan rows together with its byte accounting and done flag.

ScanBounds

Byte-ordered scan bounds as an inclusive/exclusive pair.

ScanCursor

Forward cursor tied to one protected view. No full candidate collection. v2 (B7 follow-up): when the view’s current-value fast path is valid (see ReadView::native_range), one real backend cursor serves the whole scan — built lazily on the first Self::next_batch call and held across calls (native/native_checked). Otherwise falls back to v1’s ordered seek-per-row (next_key + read_value_with_limits). Each batch plus at most one bounded lookahead row is allocated. Callers must release consumed batches; shared outstanding-byte admission is P01.

ScanRow

One materialized scan row.

StorageStatistics

Durable byte accounting by key family, for growth diagnosis and operator metrics. All values are exact sums of stored key+value lengths for the family, taken from one read view; they exclude redb’s copy-on-write slack (compare against Store::storage_file_bytes).

Store

Ordered key/value store shared by every node: admission, deterministic expansion, group-commit publication, multi-version reads, and retention.

TempPath

A temporary store file or directory; see the module docs.

TxAccumulator

Accumulated state of one in-flight cross-service transaction. The buffer is the transaction’s uncommitted write overlay (read-buffer-first like Txn); deps records the public reads performed inside the transaction; mutations/preconditions replay the writes atomically at commit.

Txn

Buffered view handed to admission/apply: reads see committed data merged with staged writes; writes stay buffered until the closure succeeds.

VersionEntry

One superseded value of a key, covering the half-open range [from,to).

WriterHolds

One holder’s writer holds since the store opened.

Enums

CommitOutcome

Result of a commit attempt: either the durable commit index plus the caller payload, or a conflict reason.

CommitPhase

One timed step of the commit pipeline. Committer-side phases are observed once per commit; publisher-side phases once per publish group.

CompactOutcome

What Store::compact_if_needed did, so a deferral is never silent.

CopyOnWrite

Copy-on-write state of a store’s data files, reported at boot.

Op

Operation emitted into the atomic publish set.

OutcomeKind

Terminal status of a commit key, as stored durably.

ScanProjection

Which columns a scan should materialize for each row.

StorageError

Errors returned by the storage layer: failed preconditions, corrupt or unsupported durable state, resource exhaustion, and invalid read requests.

WriterHolder

What holds a backend’s single writer: one redb write transaction, from begin_write to its commit. Every other writer waits meanwhile, so the longest holds are what stall other tenants’ commits.

Constants

SERIALIZABLE_CONFLICT

Reason prefix of a CommitOutcome::Conflicted caused by a read dependency that a later commit wrote (first-committer-wins).

WRITER_HOLD_BOUNDS

Upper bounds, in seconds, of the writer-hold histogram buckets.

Traits

DeriveRead

Read access to the state before a commit, for a Deriver.

Deriver

Deterministic derivation of derived rows from a commit’s base ops.

Replicator

Replication seam: when installed on a Store, local commits build the canonical CommitBatch and route it through the leader. The leader replicates to the quorum and every node (leader included) applies it through Store::apply_replicated_batch, which returns after the batch is durably applied locally. Serialized payload is stored with the outcome so idempotent retries can reproduce the caller’s response.

Functions

is_usage_bytes_key

Whether key is a tenant’s usage-bytes row.

prefix_range

Inclusive start / exclusive end range covering all keys under a prefix. An empty end bound means unbounded above.

remove_store

Remove the materialization file at path. The WAL directory and layout manifest are left alone.

Type Aliases

CommitObserver

Callback receiving each timed commit phase, in seconds.

ExpandFn

Deterministic state-machine expansion callback: derives every canonical and derived key implied by one committed batch (document 07).