How PlanetScale Powers the Mixedbread Control Plane

Mixedbread is retrieval infrastructure for AI agents. We help developers retrieve the right evidence across text, PDFs, tables, images, audio, and video with high accuracy and low latency.

Every week, Mixedbread processes millions of document and search operations across the platform. Behind every activity is a critical system: the control plane that governs access, resource lifecycle, operational coordination, and usage accounting.

The State Behind Every SearchLink to section

Mixedbread's retrieval data plane is S3-native. Documents and indexes are persisted in object storage, then loaded into faster compute and storage tiers for indexing and search. With Bring Your Own Bucket, customers can keep this data in object storage they own and control.

Separate from the retrieval data is the operational state required to run the platform safely and smoothly. PlanetScale provides the durable source of truth for four critical control-plane responsibilities:

• Access: Organization membership, credentials, and fine-grained store permissions. Every ingestion and search operation must respect these policies to prevent unauthorized access or unintended changes.

• Resource Lifecycle: How stores are created, configured, updated, and retired. This state determines how ingestion and search should behave throughout the life of a store.

• Operational Coordination: Durable state for work moving through Mixedbread's distributed ingestion and processing systems.

• Usage Accounting: Quotas, consumption, and billing-related state required to meter usage correctly.

The control plane must be fast enough to avoid impeding retrieval performance. It must provide a consistent source of truth so that access and lifecycle changes take effect correctly across the system. And it must be reliable because these decisions determine whether work can safely proceed.

Why We Chose PlanetScale MetalLink to section

Before PlanetScale, we operated our own Postgres clusters. As data demand continued to grow, the team faced growing operational overhead from managing them, including replication, backups, failovers, tuning, and networking. Meanwhile, we wanted to spend more of our time scaling our retrieval engine and building features that help our users succeed. We decided to look for a managed solution that would help us refocus on what matters.

We had followed PlanetScale's database engineering work and learned from its technical writing for years. When PlanetScale for Postgres became generally available in September 2025, it was a natural choice for us to evaluate. We began moving our control plane shortly afterward.

We chose PlanetScale Metal from the beginning because we wanted performance headroom for a critical part of our infrastructure. Metal uses locally attached NVMe storage rather than relying on network-attached storage, providing more consistent I/O characteristics. For Mixedbread, it helps keep critical control-plane operations fast as our infrastructure continues to grow.

Moving to PlanetScale was therefore more than a change in where Postgres runs. It transferred much of the operational responsibility to a team focused on databases, allowing our team to stay focused on retrieval.

The ResultsLink to section

Moving our control plane to PlanetScale delivered predictable performance, removed a significant amount of operational work, and gave us better tools for understanding how the database behaves in production.

• Predictably low latency: PlanetScale executes our hottest control-plane queries with sub-millisecond latency. Over a recent 24-hour window, the access-control read that gates nearly every API request held approximately 0.05 ms at p99 across millions of executions, and our busiest resource-lifecycle write held approximately 0.7 ms at p99 at similar volume. Every one of our top query patterns stays under 1.5 ms at p99. This keeps the database from becoming a bottleneck in the operations surrounding ingestion and search.

A line chart of hourly p99 latency for the two busiest control-plane queries over 24 hours. The access-control read holds a flat line at about 0.05 milliseconds. The resource-lifecycle write stays between 0.55 and 1.1 milliseconds. Both are measured server-side by PlanetScale.

The two busiest queries hold p99 at 0.05 ms and 0.7 ms

PlanetScale Query Insights · server-side execution time · millions of executions per day each

A horizontal bar chart of p99 latency for the seven hottest control-plane query patterns over 24 hours. Every bar is under 1.5 milliseconds: access check 0.05, resource lifecycle update 0.71, resource lookup 0.49, resource creation 0.76, metadata read 0.10, resource registration 1.44, and credential lookup 0.04.

Every hot query pattern executes in under 1.5 ms at p99

access check

0.05 ms

resource lifecycle update

0.71 ms

resource lookup

0.49 ms

resource creation

0.76 ms

metadata read

0.10 ms

resource registration

1.44 ms

credential lookup

0.04 ms

PlanetScale Query Insights · server-side execution time · 24 hours, Aug 10–11, 2026 (UTC)

• Streamlined database operations: Scaling the database is now a configuration change. We can increase capacity and configure managed PgBouncer connection pooling without provisioning new clusters or managing the rollout ourselves. When we adjusted our pooling configuration, PlanetScale applied the changes without user-visible downtime.

• From insights to actions in minutes: The team has connected to the PlanetScale MCP server, making insights and metrics directly available to our agents. PlanetScale skills ground them in established Postgres practices. These tools help us integrate Postgres context into our agentic observability stack, giving engineers a complete view of system health. When a service slows down, agents can help trace the signal from the affected operation to the relevant query pattern and the code that produced it, substantially shortening the path from detecting a problem in production to deploying a fix.

What's NextLink to section

With PlanetScale making database operations trivial, the team can focus on delivering the most value to our users: making retrieval more accurate, reliable, and scalable.