Read summarized version with
About project
Working time:
Since 2025
Industry:
Retail
The service:
Managed Services
Overview
The customer is a European multi-brand retailer selling in several countries, in stores and online. Dedicatted built OpenSearch-based retrieval for its demand forecasting platform, and its engineering teams also use Amazon OpenSearch Service for application and security logs.
The Challenge
The domains had grown faster than anyone was looking after them.
1. Shard Pressure
Daily indices created thousands of small shards, and JVM memory pressure spiked during peak trading.
2. Red Cluster Events
A node failure turned the logging cluster red during a promotion weekend.
3. Deferred Upgrades
Engine upgrades had been postponed twice.
4. Retention Cost
Every log stayed on hot storage for the full retention period.
Technology & Approach
We took over the domains under a monthly managed service, fixed the structural causes first and then kept them healthy.
Implementation
Onboarding
Reviewed domain health, instance types, shard strategy, index templates, snapshots and fine-grained access control, and set alarms on cluster status, JVM memory pressure, storage and search latency.
Structural Fixes
Consolidated indices with rollover and Index State Management policies, right-sized shards, and moved older logs to UltraWarm and cold storage.
Monthly Operations
Our engineers respond to incidents around the clock, rebalance shards, test snapshot restores and ran a blue/green engine upgrade planned with the application teams.
Vector Care and Reporting
Reviewed vector index settings and recall for the forecasting retrieval as data grew, with a monthly health and cost report.
Customization & Client Adaptation
Retail trading peaks set the rhythm of the service.
Ready for Peak Season
Capacity and shard reviews are scheduled before each major promotion.
Retention Without Hot Storage
Logs move to UltraWarm after 7 days and to cold storage after 30, and stay searchable when needed.
Tested Restores
Snapshot restores are tested each quarter into a separate domain.
Recall Tracked for AI
Retrieval quality for the forecasting platform is measured on a fixed question set as indices grow.
If you find this case interesting, we recommend taking a closer look at
The Results
Since onboarding the domains have had no red cluster events during trading peaks, the engine upgrade went through without downtime, and tiered storage cut log storage cost by roughly 40 percent.
Long-Term Value
- Index lifecycle that keeps shard counts under control.
- Upgrades planned instead of urgent.
- Retrieval recall that holds as data grows.
Business Impact
- Search and logging stay available through promotions.
- Lower retention cost without losing history.
- Engineering teams freed from cluster firefighting.
