Why Replica Cost Creeps Back In and How to Prevent It

Last week, we covered the levers for reducing read replica cost: improving query hygiene, pooling connections, reducing lag, and eliminating redundant work. This week is about where those fixes break down in practice and how to prevent it.

• Treat query hygiene as a moving target, not a one-time cleanup: Replica load rarely comes from a single bad query. It accumulates as new endpoints, background jobs, and ORM changes quietly reintroduce waste. Teams that regularly re-audit read patterns catch regressions early and avoid scaling replicas to compensate for drift.
• Separate freshness from correctness explicitly: Many replica reads don't require real-time data; they just need consistency within a request or page load. When freshness requirements are made explicit, teams can cache, batch, or route reads safely instead of defaulting everything to "latest," which needlessly increases replica pressure.
• Audit background and internal traffic independently from user traffic: Cron jobs, health checks, async workers, and internal services often hit replicas far more frequently than user-facing paths. These workloads are easy to overlook and disproportionately expensive at scale, so they need to be reviewed and budgeted separately.

Reducing replica cost isn't about clever tricks; it's about consistently removing unnecessary work. This is why many teams use Readyset to offload repetitive reads, rather than scaling expensive replicas. Teams that treat replica load as a signal, not a fixed cost, scale further with fewer replicas, lower lag, and more predictable performance.

🔍 What We're Watching in the Database Ecosystem

OpenAI Unveils GPT-5.2

OpenAI has released GPT-5.2, its newest large language model series built for professional knowledge work, long-context reasoning, coding, and complex tool use. Early reports point to real productivity gains, with hours saved on tasks like spreadsheets, presentations, and multi-step workflows, alongside stronger long-document understanding and state-of-the-art benchmark performance.

Databricks raises $4B at $134B valuation as its AI business heats up

Databricks is reportedly in talks to raise around $4 billion at a roughly $134 billion valuation, reflecting strong investor confidence in its role as a core enterprise AI and data platform. The company has repeatedly increased its revenue forecasts this year even as heavier AI usage pressures gross margins, underscoring both the appetite for AI infrastructure and the economic tradeoffs shaping data + AI platforms.

Google Research Advances AI Long-Term Memory with Titans + MIRAS

Google introduced the Titans architecture and MIRAS framework to enable AI models to build long-term memory by updating internal state as data streams in. This test-time learning approach targets the limits of fixed context windows, allowing models to reason across millions of tokens without losing critical information.

In Case You Missed It

As we head into the holiday season and the new year, we want to thank you for reading and wish you a happy holiday!

We will be kicking off 2026 at FOSDEM, where we will be hosting a happy hour (details coming soon) and co-sponsoring the Belgian MySQL, MariaDB, and Friends community dinner on Friday, January 30 (registration here). If you are attending or will be in the area, we would love to connect.