Skip to content

Glossary

Short definitions for terms used across Consist docs. For the mental model, see Core Concepts.

Term Meaning
Artifact A tracked file or file-like output with path, format, fingerprint, direction, and producing or consuming run metadata.
Cache hit A run lookup result where Consist finds a prior completed run with the same signature and returns its output artifact metadata instead of executing again.
Cache miss A run lookup result where no matching completed run exists, so Consist executes the step and records a new run.
Canonical hashing Deterministic hashing of config, inputs, and identity metadata so equivalent values produce the same signature.
Config Run parameters included in the signature through config={...}. Changing config invalidates the cache for that run.
Coupler The scenario helper that maps named outputs from one step into named inputs for later steps.
DLT Optional data loading integration used by Consist to ingest typed tabular data into DuckDB with provenance columns.
Derivable references Maintenance condition where Consist can infer safe table relationships from stored keys such as run_id and content_hash.
Facet Queryable run metadata, such as year, scenario, stage, or phase, used to find and compare runs.
Ghost mode Recovery mode where provenance or ingested data remains queryable even when the original cold file is missing.
Hydration Restoring artifact metadata and usable artifact objects from prior runs without necessarily copying bytes.
Config identity The configuration component (full config dict or adapter-derived structure) that contributes to a run signature.
DuckDB ingestion Loading artifact contents into DuckDB for SQL-native analysis and recovery.
Lineage The dependency chain from an output artifact back through the runs, inputs, configs, and upstream artifacts that produced it.
Filesystem materialization Ensuring bytes exist at a requested filesystem destination, either by copying historical files or exporting ingested data.
Provenance Evidence for what produced a result: code, config, inputs, environment metadata, outputs, and status.
Recovery roots Ordered archive or mirror roots stored on artifact metadata so future hydration can find bytes after the original workspace path moves.
Run A single tracked execution, cache hit, or trace block with inputs, config, outputs, status, timing, and metadata.
Run-link table A global table containing run_id but not consist_run_id; used for run-associated metadata and safe maintenance derivation.
Run-scoped table A global table containing consist_run_id; rows are directly attributable to a specific run.
Scenario A parent run that groups related child runs for a multi-step workflow, iteration, or scenario variant.
Signature The deterministic fingerprint built from code identity, config identity, inputs, and selected execution metadata. Used for cache lookup.
Staging Copying a canonical artifact to a requested local path for a path-bound callable or external tool.
Trace An always-execute tracked block that records provenance without cache-skipping the block.
Unscoped cache table A global table with neither run_id nor consist_run_id; preserved by default unless maintenance can prove safe pruning.
View A registered SQLModel-backed query surface for analyzing runs or artifact data.
Virtualization Querying file-backed and/or ingested artifacts through DuckDB without eagerly loading every file into memory.
Hybrid view A SQL view that combines hot ingested data with cold file-backed data so runs can be queried together.

See also: Caching & Hydration, Data Storage and Ingestion, and Historical Recovery.