The Composable CDP Decision Memo
Signals you've outgrown your packaged CDP, how to sequence a zero-copy migration to the lakehouse, and the hidden costs vendors never quote. A decision memo for 2026.
On this page
Executive summary
To: CMO, CDO, VP Engineering. Re: whether to renew the packaged CDP or migrate activation onto the lakehouse.
Our recommendation, stated up front: if your customer data already lives in Snowflake, Databricks, or BigQuery, and your CDP now functions mostly as an expensive second copy of that data, begin the unbundle and time the cutover to your renewal date. If your warehouse practice is immature, or the CDP's identity resolution is genuinely load-bearing, a hybrid zero-copy arrangement buys you 12 to 18 months without a rebuild. Doing nothing is the only option we cannot defend, because the vendors themselves have already conceded the architecture.
Consider what the incumbents shipped in a single month. Salesforce released Zero Copy File Federation for Data Cloud in June 2025, reading Apache Iceberg tables in place from Databricks, Snowflake, and AWS Lake Formation. That same month, Oracle wired zero-copy links between its Unity CDP and its Cloud Infrastructure lakehouse. When platform vendors rearchitect their flagship products around not copying your data, the case for continuing to copy it has effectively closed.
The evidence, briefly
Three facts frame this decision.
First, the money is real and the waste is documented. Grand View Research sizes the CDP market at $8.26 billion in 2025, projecting $58.41 billion by 2033 at a 27.8% compound annual growth rate. Yet Gartner's marketing technology surveys keep finding that roughly half of purchased martech capability goes unused, and only a small minority of organizations qualify as high performers meeting strategic goals with positive ROI. The packaged CDP is frequently the single largest line item inside that underused stack. Spend is accelerating into a category most buyers cannot fully operate.
Second, the substrate has standardized. The 2025 State of the Apache Iceberg Ecosystem survey, published by Data Lakehouse Hub, found 78.6% of respondents using Iceberg exclusively among open table formats, with Delta Lake overlapping at 39.3%. This matters more than any vendor announcement. Open table formats mean the same physical files can be read by your warehouse, your activation layer, your BI tool, and your ML pipelines simultaneously. Zero-copy stops being a sales slogan and becomes a property of the storage layer itself.
Third, the category's own growth pattern is telling. The CDP Institute's July 2025 industry update showed the market's strongest organic growth since 2022, a 3.4% employment increase across existing firms, while adding almost no new entrants. Mature categories consolidate. They also get unbundled from below, which is precisely the thesis Andreessen Horowitz laid out in its warehouse-first argument for the composable CDP: minimize replication, compose best-of-breed pieces over the data cloud you already run.
Signals you have outgrown the packaged CDP
No single symptom justifies a migration. Three or more of the following, observed over two consecutive quarters, do.
- Duplicated logic. Your analytics team rebuilds CDP segments in dbt because marketing stopped trusting the CDP's counts. Two definitions of "active customer" now compete in every QBR.
- Profile drift. The CDP's unified profile count and your warehouse's customer count disagree by more than 10%, and nobody can produce a reconciliation.
- Paying twice for the same bytes. Your contract prices by events or profiles, meaning you pay the vendor to store and process data you already store and process in the lakehouse.
- The connector queue. New sources land in the warehouse within a day through your existing pipelines, while the same data waits weeks for a vendor connector or professional-services ticket.
- Exported intelligence. Churn scores, predicted LTV, and propensity models are computed in the lakehouse, then shipped into the CDP through brittle CSV uploads or rate-limited APIs. If this is you, run the numbers on what those scores are worth with our CAC and LTV calculator before deciding where they should live.
- Governance ambiguity. Legal asks where consent is enforced and the honest answer is "in two places, differently."
Most of these reduce to one underlying problem: the CDP holds a slowly diverging copy of the truth. We wrote about the organizational fix in Building a Data Activation Framework That Drives Business Results; the architectural fix is what this memo covers.
What zero-copy actually changes
A packaged CDP bundles four functions over its own proprietary store: ingestion, identity resolution, segmentation, and activation. The composable pattern reassigns each of them. Storage and compute stay in the lakehouse. Identity resolution runs as warehouse-native jobs or a specialist entity-resolution service. Segmentation becomes SQL and dbt models your analysts already maintain. Activation is handled by reverse ETL or by the destination platforms reading shared tables directly.
Zero-copy is the connective tissue. Instead of nightly batch exports into a vendor silo, the activation layer queries Iceberg or Delta tables in place. One copy of the truth, many readers. The practical consequences are larger than the architectural diagram suggests: audience definitions update when the model updates, enrichment happens once in the warehouse rather than per-tool, and the data team debugging a bad audience reads the same tables the marketer built it from. The same single-copy foundation is what makes warehouse-side enrichment compound, a pattern we detailed in AI Audience Segmentation in SaaS: The Data Enrichment Playbook and its ecommerce counterpart.
One caution. Zero-copy federation is priced by the platforms too, often in compute credits per row scanned, and a poorly partitioned table queried hourly can cost more than the batch export it replaced. Read the pricing appendix before celebrating the architecture.
The hidden costs nobody quotes
Every composable vendor's pitch deck compares its license fee to the packaged CDP's license fee. That comparison flatters the composable side because the composable model moves cost from license to infrastructure and headcount. Ranges below reflect what we see in EGGKNITE migration engagements at mid-market and enterprise scale.
| Cost | Why it never appears in the sales cycle | Typical range |
|---|---|---|
| Warehouse compute for segmentation and identity | Vendors quote licenses; audience queries and match jobs run on your Snowflake or Databricks bill | 10–30% added to current warehouse spend |
| Reverse ETL pricing at scale | Row-based pricing looks trivial in the pilot and compounds with sync frequency | $12k–$60k per year at maturity |
| Identity resolution rebuild | Packaged CDPs bundle deterministic and probabilistic matching you must now own or buy | 1–3 months of data engineering |
| Dual-running both systems | You pay both platforms until audience parity is verified | Typically two renewal quarters |
| Real-time gaps | Lakehouse batch cadence cannot serve sub-second onsite decisioning alone | A slim event layer, often retained |
| Governance ownership | Consent enforcement shifts from vendor SLA to your own policies and tests | 0.5–1 FTE, ongoing |
The dual-running line deserves emphasis. Teams consistently underestimate it because they plan the migration as an engineering project and forget it is also a procurement project. If your renewal lands in month four and parity lands in month nine, you will sign one more annual contract than you budgeted for.
Sequencing the unbundle
Order the work by identity complexity, lowest first.
Paid audience syncs move first because they tolerate imperfection: a suppression list or lookalike seed built from warehouse tables and pushed to Meta and Google through reverse ETL can run in parallel with the CDP's version, and the ad platforms themselves will tell you which performs. Lifecycle email and SMS move second, once your warehouse-side segment definitions have survived a quarter of scrutiny. Onsite personalization and any sub-second decisioning move last, and honestly may keep a thin real-time layer indefinitely; forcing batch infrastructure to do streaming work is how composable migrations acquire a bad reputation.
Identity resolution is the final boss, and it should be. Run the CDP's matching and your warehouse-native matching side by side for at least one full quarter, reconcile the deltas, and only then cut over the systems that depend on a unified profile. Throughout, treat consent as code: policies expressed as tests in the warehouse, verified on every run, per the practices in our first-party data playbook.
Exit at the renewal date. Never mid-term.
Recommendation
The one-page version for your leadership team: the industry has already standardized on open tables and in-place reads, the outgrowth signals are observable in your own stack this week, and the true cost of unbundling is knowable if you price compute, reverse ETL, dual-running, and governance honestly. Decide against your renewal calendar, sequence by identity complexity, and keep exactly as much real-time infrastructure as your use cases prove they need.