How a health data platform cut GCP costs by 80% in six weeks

·1 views

A federated health data platform operating across two districts in Indonesia saw its Google Cloud Platform bill triple between March and May 2026, prompting an urgent cost audit. The platform, which converts data from 12 source systems into FHIR R4 resources, was processing over 500,000 records per day across 80 source tables. Large upstream data backfills — one pushing 2.27 million rows in a single week — caused the main FHIR resource table to grow from 150 GiB to 201 GiB. A deduplication scheduler scanning the full datastore three times per hour became the system's costliest query as data volume grew, with daily BigQuery scan volume jumping from 2.36 TiB to over 9 TiB within six weeks. The team used billing exports and BigQuery's INFORMATION_SCHEMA to pinpoint the source of the spike before making targeted infrastructure changes that ultimately reduced daily costs by 80%.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

What ADHD Paralysis Is and Why Willpower Alone Cannot Fix It

ADHD paralysis is a state in which an overload of tasks, choices, or information overwhelms executive function, leaving a person unable to initiate any action at all. Unlike procrastination, which involves avoiding a specific task, paralysis means the brain's prioritization system shuts down entirely when demands pile up. Researcher Dr. Russell Barkley attributes this to deficits in working memory and prioritization rooted in how the ADHD brain processes dopamine and norepinephrine. A 2023 study in the Journal of Attention Disorders found that adults with ADHD experience task-induced freezing at significantly higher rates than neurotypical adults, with the freeze strongly linked to anxiety and perfectionism rather than task difficulty. Experts suggest the solution lies in reducing cognitive load on the executive function system, not in applying more willpower.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How Production MCP Servers Break and the Rules One Team Built to Fix Them

A development team running two production MCP servers — an analytics gateway and an AI coding tool called AI-Lens — documented recurring failures that only emerged after real-world deployment, not during demos. Key issues included unhelpful error messages that caused AI models to hallucinate incorrect values, users being unexpectedly logged out mid-session due to refresh token rotation conflicts, and tools silently disappearing from clients. The team found that structuring error responses with actionable hints — such as listing valid table names or suggesting query optimizations — dramatically reduced wasted retries by giving the model enough context to self-correct. They also identified a critical timeout mismatch, where client and server timeouts aligned too closely, preventing structured errors from reaching the model before the connection dropped. These findings were codified into an open standard published on GitHub under a CC BY 4.0 license, complete with a release-gate checklist for MCP server deployments.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How chronyc Replaces ntpdate for Time Sync on Anolis OS 8.9 GBase Clusters

GBase 8a database clusters running on Anolis OS 8.9 can experience significant clock drift after extended downtime. Since Anolis OS 8.9 no longer includes ntpdate or the legacy NTP service, administrators must use chronyc for time synchronization. The chronyc makestep command forces an immediate clock correction, while chronyc tracking allows users to verify sync status by checking the system time offset and leap status. In a documented example, the system clock jumped nearly nine hours to align with the NTP server after running makestep. Once synchronized, the system time offset dropped to near zero, confirming successful time correction.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Five Memory Types That Make TypeScript AI Agents More Reliable

A technical guide published on DEV Community outlines five distinct memory types that developers should implement when building AI agents in TypeScript. The author argues that most agent failures — such as repeated mistakes, lost context, or inconsistent behavior — stem from poor memory design rather than weak models or prompts. The five layers covered are short-term, semantic, episodic, procedural, and audit memory, each serving a different purpose in the agent's decision-making process. While short-term memory tracks the current run and episodic memory logs past events, procedural memory enforces rules and audit memory supports compliance and debugging. The piece concludes that structuring memory correctly is the most effective way to make AI agents predictable and production-ready.

0 comments Read more at DEV Community