Selected work · Zillow
Cut the bill without cutting the platform
ETL cost was rising faster than the value it produced, across stacks that had each been a reasonable local decision. Consolidating science, MLOps, and engineering onto one lakehouse path cut ETL cost 82 percent and made queries 43 percent faster, without a freeze on the roadmap.
- 82%
- ETL cost reduction
- 43%
- faster queries
- 21%
- less engineering overhead
The situation
Data science, MLOps, and data engineering had each built the stack that suited them. Every one of those choices had been defensible on the day it was made. Together they meant the same data was moved, stored, and reprocessed several times over, and the bill grew in proportion to the duplication rather than to the business.
The pressure showed up as a finance question, but it was an architecture question wearing a finance costume. Nobody could model the bill, so nobody could argue about it with evidence, so the conversation defaulted to across-the-board cuts that would have slowed product and ML.
Constraints
- No roadmap freeze. Product and ML work had to keep shipping through the consolidation.
- Ownership was moving outward at the same time, toward the teams producing the data. Controls had to get stronger while the org got more distributed, not weaker.
- Real-time clickstream was becoming a product requirement, so the target had to support streaming rather than being a cheaper batch stack.
What I did
01
Model the bill before touching the architecture
Cost per pipeline, per team, and per unit of business value, so the expensive paths could be named. Most of the spend was concentrated in a small number of jobs. That is almost always true, and it is almost never known before someone does the arithmetic.
02
Consolidate onto one plane, in dependency order
Data science, MLOps, and engineering moved onto a shared lakehouse path, sequenced so each step paid for the next. Migrations that ask for the full budget before returning anything tend to get cancelled halfway, which leaves you running both stacks.
03
Decentralize ownership and tighten governance together
Ownership moved to producing teams while controls got stricter, not looser. Decentralization without governance is how you end up with the fragmentation you just paid to remove.
04
Put real-time on the decision loop
Clickstream moved onto the product decision path in real time, so the consolidation delivered a new capability rather than only a smaller invoice. A cost program that produces no new capability is a hard sell the second time you need one.
Outcomes
- One plane for science, MLOps, and engineering
- Clearer ownership without weaker controls
- Clickstream feeding product decisions in real time
- Cost outcomes leadership could take to finance
What I would tell you now
- Cost problems are usually architecture problems that reached finance first.
- Sequence a migration so each phase returns something. Self-funding beats a big-bang plan that needs unbroken faith for a year.
- The number that matters is not the percentage saved, it is whether leadership can explain the saving to finance without you in the room.
Who this is for
Read this if your cloud bill is outgrowing the business, if several teams have each built a reasonable stack that adds up to an unreasonable one, or if you are being asked to cut spend without slowing delivery.
Recognize the shape of this?
Thirty minutes is usually enough to tell whether it is the same problem. I reply within 48 hours.