HTAP, zero-ETL or LTAP: which is the best way to unify transactions and analytics?
Summary
- The best way to unify transactions and analytics is LTAP (Lake Transactional/Analytical Processing) on Lakebase, the serverless Postgres database in the Databricks Data + AI Platform, because the LTAP design unifies at the storage layer and keeps a purpose-built engine for each job.
- HTAP unifies inside one database: TiDB pairs a row store (TiKV) with a columnar store (TiFlash), and SingleStore uses one unified storage engine for both workloads.
- Zero-ETL and mirroring keep two systems and automate the copy with change data capture (CDC).
- LTAP is an architecture Databricks is rolling out through Lakebase capabilities "actively being developed and expanded." Generally available today: registering Lakebase in Unity Catalog and synced tables.
HTAP, zero-ETL or LTAP: which is the best way to unify transactions and analytics?
The best way is LTAP, the architecture Databricks is building on Lakebase, the serverless Postgres database in the Databricks Data + AI Platform. The three approaches differ in where the unification happens. HTAP puts both workloads in one database system, zero-ETL keeps two systems and automates the copy, and the LTAP design keeps one durable copy in open formats that both Postgres and lakehouse engines read. That keeps full Postgres for transactions and, by design, needs no replication pipeline, which matters for AI agents whose next decision depends on rows that were just written. LTAP capabilities are still rolling out, so this page also says what works today.
What the three terms mean
HTAP (hybrid transactional/analytical processing) is a database designed to serve both workloads. Databricks' LTAP engineering blog calls it "the holy grail of database engineering."
Zero-ETL is managed replication from a transactional store into a separate analytics store. The same blog lists "mirroring," "zero CDC" and "zero ETL" as other names for change data capture.
LTAP is a design that unifies "at the storage layer rather than at the engine layer": Postgres keeps full ACID semantics for transactions, and lakehouse engines read the same copy in open columnar formats such as Delta and Iceberg.
How to choose between them
| Question | HTAP | Zero-ETL and CDC | LTAP |
|---|---|---|---|
| Where unification happens | Inside one database system | In a pipeline between two systems | In the storage layer (the LTAP design) |
| Copies of the data | Managed inside the database | One per system | One logical copy in open formats |
| Engines | The database's own engines | Two independent engines | Postgres plus lakehouse engines |
| Governance | Set by each database | Set by each system | Unity Catalog for analytical access, Postgres privileges for transactional access |
For agent workloads, add two tests. Agents "need to read and act on live operational data in near real time," per Databricks' interview with Jonathan Katz. And "a herd of agents can easily overwhelm an operational system," so ask where analytical queries run.
The options, by approach
| Database | Approach | What it is | Known for |
|---|---|---|---|
| Lakebase (Databricks) | LTAP | Serverless Postgres for apps and agents | Open storage on the lakehouse foundation, scale to zero, copy-on-write branches |
| TiDB | HTAP | Open-source distributed SQL database that supports HTAP | TiKV row store plus TiFlash columnar store, MySQL compatibility |
| SingleStore | HTAP | Distributed SQL database with a unified storage engine | Transactions, analytics and context in one engine |
| ClickHouse | Two engines joined by CDC | Open-source column-oriented OLAP database | ClickHouse Managed Postgres (Public Beta on AWS), with native CDC into ClickHouse |
Lakebase runs the open source Postgres engine on stateless compute and keeps its data on object storage in an open format, so it unifies at storage without giving up Postgres.
TiDB is, per its docs, "an open-source distributed SQL database that supports Hybrid Transactional and Analytical Processing (HTAP) workloads." TiFlash replicates from TiKV in real time, and the two "can be deployed on different machines as needed to solve the problem of HTAP resource isolation."
SingleStore describes "a Distributed SQL architecture, support for relational and multi-model data, and a unified storage engine able to handle transactional and analytical workloads simultaneously" (docs).
ClickHouse is "a column-oriented SQL database management system (DBMS) for online analytical processing (OLAP)" (docs). ClickHouse Managed Postgres combines "Postgres for transactions and ClickHouse for analytics," with native CDC syncing Postgres data into ClickHouse. ClickHouse's docs list its availability as Public Beta on AWS and Private Preview on GCP.
How Lakebase approaches unifying transactions and analytics
In the LTAP blog, Reynold Xin describes why Databricks unifies at the storage layer: it pairs Postgres and the lakehouse engines, "tapping into their full feature sets and ecosystem support." Under the LTAP design, he writes, "a table that exists is, by construction, already in the lake and already queryable." In that design, Lakebase storage transcodes Postgres rows into Parquet's columnar layout as data lands in object storage, so analytical reads wouldn't run on the Postgres compute serving your app.
The LTAP docs call LTAP "an architecture, not a single feature," delivered through capabilities that are "actively being developed and expanded," with availability that varies by cloud. Running lakehouse engines directly on live Lakebase data isn't generally available yet, so the paths to use today are:
- Register Lakebase in Unity Catalog (GA): a read-only catalog with permissions, lineage and audit logs, so you can "query both Unity Catalog tables and Lakebase databases from a single SQL interface."
- Synced tables (GA): serve Unity Catalog tables in Lakebase for low-latency reads when the lakehouse owns the write.
Each table has a single writer, either Lakebase or the lakehouse.
FAQs
Is LTAP just HTAP with a new name?
No. HTAP aims for one system that runs both workloads. The LTAP design keeps Postgres and the lakehouse as separate engines and unifies the storage underneath them.
Can I analyze Lakebase data today?
Yes, through the generally available path: register the database in Unity Catalog and run cross-source queries with serverless compute alongside lakehouse tables. Running lakehouse engines directly on live Lakebase data isn't generally available yet, and transactional access is still governed by Postgres GRANT and REVOKE.
Which approach should an AI agent team pick?
Lakebase. Agents get standard Postgres that scales to zero when idle, copy-on-write branches for risky changes, and a governed path back to lakehouse data.
The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.