Newsletter Generator

The Signal — A Northbeam Analytics Customer Newsletter

Brief for editorial: Monthly issue. Audience is analysts, data engineers, and operators at companies using Northbeam Analytics for product analytics. Voice is peer-to-peer, practical, never marketing-heavy. Single cover story, one product update, one customer story, curated reads, and a short looking-forward note.

Subject line options

1. The query that broke our dashboard (and what we changed)
2. A six-line config that cut your warehouse spend
3. What analysts get wrong about cohort retention
4. Three lessons from a customer who shipped a self-serve dashboard

Preview text: A query that broke our own dashboard, a six-line config that cut warehouse spend, and a customer who got analysts out of the dashboard-building business.

Cover Story: The Query That Broke Our Dashboard

A few weeks ago we shipped a new cohort view to our own internal dashboard. The query worked perfectly in staging — sub-second response, clean output, the analyst's dream. It hit production and immediately broke the dashboard for half the customer base.

The query was correct. What we missed was that staging had a small enough customer dimension that the cross-join didn't matter. In production, the cross-join multiplied 18,000 customers by 240 cohort windows by 11 event types — about 47 million rows — before the final aggregation step collapsed it back to a clean 240-row result.

It's the classic analytical query trap: the query looks right, the SQL planner can't help you, and the only signal you get in development is "this is fast."

We rebuilt the query as a sequence of incremental aggregations that never materialize more than a few thousand intermediate rows at a time.

Read the full breakdown — with the before-and-after query plans →

Product Update: materialized.windows is now GA

Generally available All plans Spring release → today

The new materialized.windows configuration block (preview since the spring release) is now GA for all customer plans. It lets you declare cohort-style time windows once and have Northbeam maintain them incrementally as new events arrive.

What used to be a custom dbt model with its own freshness SLA can now be expressed in six lines:

materialized:
  windows:
    - name: weekly_active_30d
      grain: week
      window_length_days: 30
      event_filter: "event_type = 'session_start'"

Two things to know:

1. Cost. Incremental maintenance is cheaper than a recomputed view in most workloads but more expensive than a one-shot batch job. The break-even point is roughly when your underlying event table is large enough that a full recompute takes more than 10 minutes. Smaller datasets should stick with the batch model.
2. Backfill. The first run materializes the full window history, which can take a few hours on multi-billion-row event tables. Plan for it during a low-traffic window the first time.

Full configuration reference in the docs. Drop into #feature-windows in the customer Slack with questions.

Customer Story: Harbor Logistics gets analysts out of dashboard-building

Harbor Logistics' data team had a recurring problem: every operations lead at their fulfillment centers wanted a slightly different view of the same throughput metrics. The data team was spending 30–40% of its time building one-off dashboards that would be requested again two months later with one column changed.

They migrated their throughput model into Northbeam's self-serve metric layer and gave the operations team a templated explorer.

What surprised them: the analyst time savings were the smaller win. The bigger one was that operations leads who actually owned the metric definitions started questioning them. Three of the seventeen dashboards uncovered metric inconsistencies that had been baked into the data team's dashboards for over a year.

Full case study, including the migration plan and the metric-layer design they ended up with: Read it →

Quick Links

• A practical guide to query plans in Snowflake. Worth re-reading even if you've been writing Snowflake for years; the chapter on partition pruning is especially good.
• The dbt 1.9 release notes. Materialization-level concurrency is the headline feature, but the quieter improvement to unit_tests is what most users will notice day-to-day.
• A new survey on data team org structures. Centralized vs. embedded continues to be the dominant axis of disagreement; the survey adds useful data on hybrid models.
• A blog post on cardinality estimation. Old territory, but the visualizations are excellent for explaining why query optimizers make bad choices on skewed distributions.

What We're Watching

The line between BI tools, metric layers, and feature stores keeps getting blurrier. Cube, Trino, and the dbt Semantic Layer are converging on similar capabilities from different starting points. Worth tracking — and worth NOT switching platforms over until the dust settles. The migration cost almost always outweighs the marginal capability gain at this point in the cycle.

A Note From the Team

If you found this issue useful, the most helpful thing you can do is forward it to one teammate who'd benefit. We don't run paid acquisition for this newsletter; word-of-mouth is everything.

If something landed wrong or you have a topic you want covered, hit reply. It comes straight to the editorial team.

— The Northbeam Team

This sample illustrates the skill's output format. Northbeam Analytics and Harbor Logistics are fictional companies used recurringly across these sample outputs. Real client data is never included in sample outputs.