TimescaleDB handles time-series from the last 1–90 days (hot) excellently, but at 5+ years of history the cost of PostgreSQL SSD and RAM grows faster than the business value. Apache Iceberg (a table format for the data lake) plus DuckDB (a local SQL engine) deliver a cheaper warm tier (90–730 days) and cold tier (>2 years) on S3 or MinIO. A 3-tier architecture cuts storage cost by 80–95% while keeping data queryable. This article shows how to build it for MES, where the limits are, what it costs, and how to migrate from a monolithic TimescaleDB.

OmniMES uses TimescaleDB — the PostgreSQL extension with hypertables, native compression and continuous aggregates — as its database for industrial sensor time-series. The architecture handles 200 million readings per day, delivers 95% storage compression and 200–800 ms aggregations. This article walks through the hypertables concept, the OmniMES stack architecture, continuous aggregates and concrete production numbers.

Two years after the first pretrained transformers for time series — TimesFM from Google, Chronos-Bolt from Amazon, Moirai-MoE from Salesforce — we finally have real answers to the question of whether a custom XGBoost pipeline for failure prediction in MES can be replaced by a zero-shot foundation model. This article walks through it without hype: what these models were trained on, the quality they deliver on real compressor sensor data, latencies on Jetson Orin and on a GPU server, when zero-shot is enough, and when fine-tuning is unavoidable — and what this means for the classic six-month ML project in a factory.

A comparison of InfluxDB and MongoDB for industrial data collection. This article discusses the use of InfluxDB for time-series data analysis, machine monitoring, and IoT, as well as MongoDB as a universal database for MES systems. Practical tips on when to choose each database.