Field notes from our engagements: how we build data platforms, content platforms and the AI foundations beneath them — and what actually holds up in production.
Most of an ingestion run is waiting: for late files, for Glue, for a person to fix a bad file. One checkpointed Python function per batch now does all of it, started from S3 exactly once.
Read the article →Same table, same query: 1.7 GB read from CSV, 3.8 MB from Parquet. We measured it, published the script, and show why sort order decides most of the saving.
Read the article →A file share does several jobs at once, and S3 suits some far better than others. Where each job goes on AWS, what the new S3 Files does and doesn't cover, and pipelines that start when a file is actually finished.
Read the article →Every SSIS estate has one package nobody will touch. Inventory, pattern mapping, the behavior differences that break numbers, and reconciliation that proves each move before anyone switches.
Read the article →Being AI-ready is not about connecting an app to a model. It is about the data, APIs, semantics, provenance and authorization underneath — and where we are investing next.
Read the article →Plain Parquet can't change a row without rewriting a partition. Replicating database changes into Apache Iceberg with DMS and Glue: change ordering, a rerun-safe MERGE, history, schema drift and deletes that really delete.
Read the article →The CMS you modeled for your website can be the most reliable knowledge source an AI assistant has — if it's structured. How to index it safely, keep it fresh and make every answer cite its source.
Read the article →CMS migrations rarely fail on launch day. They fail six weeks later, in the traffic report. The playbook: a URL inventory, a repeatable migration script, one-hop redirects and honest 410s.
Read the article →Most Contentful projects that disappoint fail in the content model, months before launch. Model meaning instead of layout, keep references shallow, and change the model safely once you're live.
Read the article →A one-word typo took eleven days to fix. That is an architecture problem, not a people problem — and headless is how you hand publishing back to the team that owns the message.
Read the article →Pull terabyte-scale vendor files straight into S3 without a Spark cluster — directory listing, connector transfers, the exact EventBridge event contract, and Step Functions hand-off.
Read the article →Where custom logic actually runs — DynamicFrame transforms, Glue Studio nodes, Python UDFs, pandas UDFs and applyInPandas — with the performance trade-offs and a decision tree.
Read the article →A field guide to serverless Spark at terabyte scale: storage layout, job bookmarks, shuffle tuning, skew handling, worker sizing and the cost levers that actually matter.
Read the article →They are not competitors — one orchestrates, one computes. A dimension-by-dimension comparison of Glue Workflows, Airflow and Step Functions, with a decision framework.
Read the article →Start with a free data platform assessment — we'll map your pipelines, surface quick wins, and give you a 90-day engineering roadmap.
Book an assessment
