​
High Ingest Workloads

ParadeDB is designed for real-time, high-ingest workloads. The BM25 index uses a write-optimized, log-structured architecture that minimizes contention and allows bulk writes to proceed with low latency.

As with any Postgres index, adding a BM25 index introduces additional overhead during INSERT/UPDATE operations. Every change to a row must also be reflected in the index, which involves serializing the indexed columns, tokenizing text, updating internal data structures, and potentially triggering compactions.

We’re continuously making improvements to reduce write amplification and optimize background maintenance. See our roadmap for ongoing efforts and our performance guide for tuning Postgres’ memory and parallelism settings.

​
Aggregate Performance

ParadeDB’s custom scan is optimized not only for fast text search, but also for COUNT and “top N” queries. These queries are accelerated by ParadeDB’s columnar index.

Additionally, ParadeDB provides a set of user-defined functions (UDFs) that allow developers to execute advanced aggregates. We are currently working on a new custom scan node that can translate native SQL aggregation syntax into ParadeDB’s internal execution model. Please see the roadmap for more details.

​
Distributed Workloads

ParadeDB is designed to scale vertically on a single Postgres node with potentially many read replicas, and many production deployments comfortably operate in the 1–10TB range. The largest single ParadeDB database we’ve seen in production is 10TB.

For datasets that significantly exceed this scale, ParadeDB supports partitioned tables and can be deployed in sharded Postgres configurations. If you’re working with very large datasets, please reach out to us. We’d be happy to provide guidance and share our roadmap for future distributed query support.

​
Covering Index

The BM25 index in ParadeDB is a covering index, which means it stores all indexed columns inside a single index per table. This decision is intentional — by colocating all the relevant data, ParadeDB optimizes for fast reads and boolean conditions.

However, this design comes with some tradeoffs:

  • All columns must be defined up front at index creation time. Adding or removing columns requires a REINDEX.
  • Write amplification is higher since every write must index the values of all columns into the index structure.

​
DDL Replication

A commonly known limitation of Postgres logical replication is that DDL (Data Definition Language) statements are not replicated. This includes operations like CREATE TABLE or CREATE INDEX.

If ParadeDB is running as a logical replica of a primary Postgres, DDL statements from the primary must be executed manually on the replica. We recommend version-controlling your schema changes and applying them in a coordinated, repeatable way — either through a migration tool or deployment automation — to keep source and target databases in sync.