#Glossary and Reference
Terminology you should use precisely, and figures worth carrying. Verify anything you intend to state as a hard fact against current documentation on the day — defaults and statuses move.
#Platform terms
| Term | Definition |
|---|---|
| Control plane | Databricks-managed services: web app, APIs, job scheduler, cluster manager, UC metastore service. Orchestrates; is not where your data rests |
| Compute plane | Where work executes. Classic: your cloud account/VPC. Serverless: Databricks' account |
| Workspace | A working environment; attaches to exactly one metastore |
| Metastore | The Unity Catalog root. One per region |
| Catalog | Top level of the three-level namespace; the practical isolation unit for environment and/or domain |
| External location | Storage credential + path; the grantable unit of raw storage |
| Storage credential | A cloud IAM identity Databricks assumes to reach storage |
| Credential vending | UC issuing short-lived scoped storage access at query time, so users never hold cloud keys |
| Access mode | Compute isolation setting; determines whether UC can enforce fine-grained controls |
| Cluster policy | Constraint template on compute — instance types, auto-termination, tags, runtime |
| DBU | Databricks Unit — the consumption billing unit; rate varies by SKU and Photon |
| System tables | Governed operational tables: billing, audit, query history, lineage, compute |
| Asset Bundle (DAB) | Declarative YAML packaging of jobs, pipelines and permissions for CI/CD |
| Budget policy | Attribution and cap mechanism for serverless spend |
#Spark terms
| Term | Definition |
|---|---|
| Job / stage / task | One action / shuffle-bounded unit / one partition's work in one slot |
| Narrow vs wide dependency | One-to-one partition lineage vs many-to-many (a shuffle) |
| Catalyst | The query optimiser: parse → analyse → optimise → physical plan |
| Tungsten | Memory/CPU efficiency layer: UnsafeRow, managed memory, cache-aware algorithms, codegen |
| Whole-stage codegen | Fusing a stage's operators into one generated JVM method |
| AQE | Adaptive Query Execution — runtime re-planning at shuffle boundaries |
| DPP | Dynamic Partition Pruning — runtime pruning of fact partitions from a filtered dimension |
| CBO | Cost-based optimisation; requires fresh statistics |
| Skew | Uneven key distribution making one task dominate a stage |
| Salting | Spreading a hot key across tasks with a random suffix, fanning out the other side |
| Spill | Execution memory exhausted; intermediate data written to local disk |
| Photon | Vectorised C++ execution engine; higher DBU rate, usually lower total cost on SQL/ETL |
| Spark Connect | Client/server split sending unresolved logical plans over gRPC |
#Delta terms
| Term | Definition |
|---|---|
_delta_log |
The transaction log; the definition of which files are in the table |
| Checkpoint | Periodic Parquet rollup of log state so readers don't replay from zero |
| OCC | Optimistic concurrency control — no locks; commit wins by claiming version N+1 |
| WriteSerializable | Default isolation; permits some reordering of concurrent appends |
| Data skipping | Using per-file min/max stats to avoid opening files |
| Z-order | Multi-column co-location via a space-filling curve; full rewrite, fixed keys |
| Liquid clustering | Incremental, evolvable clustering (CLUSTER BY); the modern default |
| Deletion vector | Bitmap of logically-deleted row positions (merge-on-read) |
| CDF | Change Data Feed — row-level change records readable by version/timestamp range |
| Table feature | A discrete capability a table declares, with reader/writer requirements |
| UniForm | Iceberg/Hudi metadata generated over the same Parquet files |
| VARIANT | Open binary semi-structured type, navigable without string parsing |
| Predictive Optimization | Automatic OPTIMIZE/VACUUM/statistics maintenance on managed tables |
#Current product names
| Term | Definition |
|---|---|
| Lakeflow Connect | Managed ingestion connectors (100+) |
| Lakeflow Declarative Pipelines | Declarative pipelines (formerly Delta Live Tables) |
| Lakeflow Jobs | Orchestration (formerly Workflows) |
| Lakeflow Designer | No-code authoring producing open Spark declarative pipelines |
| Lakebridge | Free warehouse-migration assessment and SQL conversion tooling |
| Lakebase | Serverless Postgres on lake storage, UC-governed |
| LTAP | Lake Transactional/Analytical Processing — operational writes landing directly in Delta/Iceberg |
| Lakehouse//RT | Real-time SQL warehouse type powered by the Reyden engine. Beta, read-only |
| Unity AI Gateway | Runtime governance of model/agent/tool traffic |
| Agent Bricks | Managed agent build/evaluate/deploy platform |
| Omnigent | Open-source meta-harness above agent harnesses |
| Genie / Genie One / Genie Ontology | Conversational analytics; agentic cross-app interface; the context layer grounding it |
| UC Metrics / Domains / Business Glossary | The governed semantic layer |
| Delta Sharing / OpenSharing | Open live data sharing; its Linux Foundation evolution covering AI assets and Iceberg clients |
| Lakewatch | Agentic SIEM on open lakehouse format |
| ZeroBus | Push API writing events directly to Delta at low latency |
#Figures worth carrying
Heuristics (say them as heuristics):
| Target file size after compaction | ~128 MB – 1 GB |
| Shuffle partition target | Low hundreds of MB |
| Partition threshold | ≳1 TB table, ≳1 GB partitions |
| Skew investigation | max/median ≳ 5× |
| Z-order columns | ≤3–4 |
| Interactive auto-termination | 30–60 min |
| Typical config-only cost win | 15–30% |
Defaults (verify on the day):
delta.logRetentionDuration |
30 days |
delta.deletedFileRetentionDuration |
7 days |
| Default isolation level | WriteSerializable |
| Indexed columns for statistics | First 32 (configurable) |
| Checkpoint interval | Every ~10 commits |
Vendor-claimed 2026 figures — quote as claimed, and say you'd validate on the customer's workload:
| Lakehouse//RT latency | ~10 ms small datasets, sub-100 ms larger |
| Lakehouse//RT concurrency | Sub-100 ms at ~12,000 queries/sec |
| Lakehouse//RT vs dedicated serving | Up to 16× |
| Lakebase instance creation / branching | Sub-500 ms |
| ZeroBus | Sub-5-second latency, 10+ GB/s |
| Lakeflow Jobs scale | ~1.7 billion job runs/month |
| Lakeflow Connect | 100+ connectors |
| Delta Sharing recipients | 28,000+, ~33% cross-platform |
Attaching "that's their number, I'd benchmark it on your workload" to any of these raises your score. It is truth-seeking in practice, and it is what a customer needs to hear.