Replace always-on clusters with scale-to-zero compute.

Always-on streaming ate the margin.

District.ai operates a multi-terabyte-per-day real-time streaming ETL pipeline backed by 24/7 SLAs to their downstream customers. The architecture worked, but the math didn't: a Databricks Photon cluster running continuously to keep latency in budget, with 432 cores online whether traffic was at peak or trickling in.

The team had already done the obvious. Cluster autoscaling. Photon. Spot instances where SLA permitted. Each round of optimization recovered single-digit percentages, never the order-of-magnitude shift the business needed.

The problem wasn't configuration. It was that JVM-based, always-on Spark clusters carry overhead the workload doesn't need.

One Test Flight. Three pipelines. Four weeks.

District scoped a Test Flight against three of their most expensive streaming pipelines. Haevek configured Falcon-native equivalents, deployed into District's existing EKS cluster alongside their Databricks workspace, and ran apples-to-apples benchmarks for four weeks against live traffic.

• No replatforming. Delta Lake stayed. Unity Catalog stayed. Falcon read from and wrote to the same paths.
• No code rewrites for Spark logic. Falcon's PySpark-compatible surface ran the team's existing transformations as-is for the first benchmark pass.
• Optimized rebuild on pipelines #2 and #3 — the same outputs, expressed natively against Falcon's compiled pipeline model, to measure ceiling performance.
• The team trained on Falcon tooling in the same engagement, so production handoff didn't require Haevek presence.

Four weeks in, the value assessment was clean: across the three pipelines, infrastructure dropped 93%, runtime dropped 80%, and operational complexity went down with it.

From signed contract to production ROI in 8 days.

District signed the subscription based on Test Flight results. Production cutover took eight days. The subscription paid for itself in seventy-five.

Within ninety days of going live on the original three pipelines, the team had identified a fourth, fifth, and sixth workload to move — quadrupling their license footprint and locking in a multi-year agreement.

The team had been looking for a 20% improvement to justify a procurement cycle. They got 93%, and the conversation changed: instead of justifying Falcon, they were justifying which Spark workloads not to migrate.

Architecture, not configuration.

Three properties of Falcon are responsible for the result. None of them are tunings of Spark.

• Compiled pipelines. Operator graphs in Falcon compile to native Rust binaries before they run. There is no interpreted execution layer to optimize at runtime.
• No JVM, no garbage collection. Memory is managed at compile time. The pauses, spikes, and per-operation tax that the JVM imposes on long-lived streaming jobs are not present.
• Scale-to-zero by default. Falcon allocates cores to workloads on demand and releases them when the work is done. The 24/7 floor most streaming engines assume is gone.

The combination is why the comparison reads as a step change rather than an incremental improvement. Same data. Same SLAs. Different engine.