Struct TaelBackend 

pub struct TaelBackend { /* private fields */ }

Implementations

impl TaelBackend

pub fn new(data_dir: &str) -> Result<Self>

pub fn with_wal_key(data_dir: &str, wal_key: &str) -> Result<Self>

Like Self::new but with an explicit WAL namespace key — lets tests run isolated instances (the WAL key is process-global in walrus).

pub fn with_wal_key_and_sinks( data_dir: &str, wal_key: &str, sinks: Vec<Arc<dyn WalSink>>, required_acks: Option<usize>, ) -> Result<Self>

Like Self::with_wal_key but with WAL replication sinks attached: this backend runs as a leader that ships every appended record to its standbys before acking the write (docs/tael-server-scaling-ha.md §5.1). required_acks is how many standbys must confirm before a write returns (None = all = fully synchronous; Some(0) = async best-effort). With no sinks the write path is unchanged. A standby on the receiving end applies shipped records via Store::apply_framed_wal.

pub fn with_components( data_dir: &str, wal_key: &str, sinks: Vec<Arc<dyn WalSink>>, required_acks: Option<usize>, cold_backend: Option<Arc<dyn ObjectBackend>>, comments: Box<dyn CommentsStore>, ) -> Result<Self>

Full constructor used by the cloud deployment path: pass an optional cold-tier object backend (None = local filesystem) and an already-built comments store (JSONL or Postgres). The local constructors above forward here with the single-binary defaults, so their behavior is unchanged.

pub fn search_index(&self) -> Arc<SearchIndex>

The shared payload search index — handed to the ingest path so prompt/ completion text is indexed at write time (the text isn’t retained on the span itself, only its blob hashes).

pub fn compact_spans(&self, cutoff: DateTime<Utc>) -> Result<usize>

Roll spans older than cutoff out of the LSM hot tier into Parquet. Returns the number of spans compacted. Safe to call repeatedly.

pub fn compact_logs_metrics(&self, cutoff: DateTime<Utc>) -> Result<usize>

Roll aged logs/metrics out of the hot tier into Parquet. Returns the total number of records compacted across both signals.

pub fn collect_live_blob_hashes(&self) -> Result<HashSet<String>>

Collect every blob hash still referenced by a live row — LLM prompt and completion hashes on spans, and body_sha256 on logs — across hot and cold tiers. Drives blob GC (anything not here is unreferenced).

pub fn enforce_span_retention(&self, keep: DateTime<Utc>) -> Result<usize>

Drop cold partitions (spans/logs/metrics) whose date is older than keep. Returns the total number of partitions removed. (Metadata GC; payload-blob GC runs separately in the maintenance task.)

Trait Implementations

impl Store for TaelBackend

fn apply_framed_wal(&self, framed: &[u8]) -> Result<()>

Standby entrypoint: durably accept a framed WAL record shipped from a leader and bring local state up to it. Mirrors the leader’s write discipline (append → apply → consume) so the standby’s WAL, hot tier, and projection stay byte-identical and itself replayable — the basis for promotion on leader loss (§5.1).

fn insert_spans(&self, spans: &[Span]) -> Result<()>

fn insert_logs(&self, logs: &[LogRecord]) -> Result<()>

fn insert_metrics(&self, metrics: &[MetricPoint]) -> Result<()>

fn query_traces(&self, query: &TraceQuery) -> Result<Vec<Span>>

fn get_trace(&self, trace_id: &str) -> Result<Vec<Span>>

fn list_services(&self) -> Result<Vec<ServiceInfo>>

fn query_logs(&self, query: &LogQuery) -> Result<Vec<LogRecord>>

fn query_metrics(&self, query: &MetricQuery) -> Result<Vec<MetricPoint>>

fn add_comment( &self, trace_id: &str, span_id: Option<&str>, author: &str, body: &str, ) -> Result<TraceComment>

fn get_comments(&self, trace_id: &str) -> Result<Vec<TraceComment>>

fn query_summary( &self, last_seconds: i64, service: Option<&str>, ) -> Result<SummaryReport>

fn query_anomalies( &self, current_seconds: i64, baseline_seconds: i64, service: Option<&str>, ) -> Result<AnomalyReport>

fn query_correlate(&self, trace_id: &str) -> Result<Option<CorrelateReport>>

fn query_sql(&self, sql: &str) -> Result<Vec<Value>>

Read-only SQL query surface (SELECT/WITH) over the telemetry tables, returning rows as JSON objects.

fn flush(&self) -> Result<()>

Flush durable buffered state ahead of a graceful shutdown. The WAL fsync on the write path is the real durability boundary, so this is best-effort: it tightens the hot tier’s on-disk state so a restart or standby replays less WAL (§5.4 “flush the hot tier”). Default is a no-op.

fn health(&self) -> Result<()>

Readiness probe — Ok(()) when this store can serve requests. Backs the REST /readyz endpoint (docs/tael-server-scaling-ha.md §5.4). The default is Ok(()): an embedded backend that constructed successfully and holds its file locks is, by definition, ready. Backends that depend on the network (e.g. RemoteStore, FanoutStore) override this to probe their dependencies.