CONCEPT Cited by 1 source
Sequence quality dimensions¶
Definition¶
Sequence quality in a multi-tenant ML data platform is multi-dimensional — not a single SLO but a vector of four orthogonal properties that consumers (training pipelines, offline analysts, online inference services) all care about in different combinations (Source: sources/2026-05-21-pinterest-making-user-sequence-data-more-cost-efficient-faster-and-easier-to-use).
The four dimensions for user event sequences:
- Freshness — "how quickly new events and enrichments show up in the sequence."
- Completeness — "whether late-arriving events, corrections, or backfills are eventually reflected."
- Consistent enrichment — "ensuring that the same enrichments are available across streaming and batch, and that training and serving see aligned data."
- Stable schemas — "downstream consumers need schemas to be versioned and predictable, not silently changed."
"One more constraint is that this is a multi-tenant platform. It has to support many teams and models, each with different needs and lifecycles. That makes correctness, observability, and operability just as important as raw throughput or latency."
Why each dimension matters¶
Freshness¶
Online inference consumers want their morning's actions reflected in ranking now. Freshness directly drives recommendation surface quality on contextual surfaces (cf. Contextual Sequential CG's daily-refresh cadence as a freshness ceiling). Closely related to landing latency in CDC pipelines but applied to the higher-level sequence-output layer.
Completeness¶
Training consumers want sequences to be eventually correct — late events, enrichment corrections, and backfills introduce new historical coverage that should propagate. Without completeness, training data drifts away from what production traffic actually saw. See late-arriving data for the underlying problem.
Freshness and completeness pull in opposite directions:
"On one hand, they need freshness: 'I want this morning's actions reflected in ranking now.' On the other hand, they care about completeness and correctness: 'If late events show up tomorrow, I still want my sequences and training data to be right.'"
A lambda architecture is the structural answer: streaming optimises for freshness, batch optimises for completeness, and they cooperate.
Consistent enrichment¶
The streaming-vs-batch axis (same enrichments available regardless of path) is the failure mode that configuration-as-code + shared execution engine are designed to prevent. The training-vs-serving axis is online-offline discrepancy — feature value mismatch between training data and inference paths. Both manifest as "silent quality degradation": nothing is broken, but model performance regresses unpredictably. Sibling failure: embedding version skew in two-tower retrieval.
Stable schemas¶
Schemas are the API contract between the platform and its tenants. Silent schema changes break consumer pipelines, invalidate cached features, and create cross-tenant blast radius. "Versioned and predictable" requires explicit deprecation lifecycles, additive-by-default change policy, and observable rollout state.
Operational mapping¶
Pinterest dashboards each dimension explicitly (Source: sources/2026-05-21-pinterest-making-user-sequence-data-more-cost-efficient-faster-and-easier-to-use):
| Dimension | Dashboarded as |
|---|---|
| Freshness | sequence freshness + lag |
| Completeness | event + enrichment coverage (cf. concepts/feature-coverage-dashboard) |
| Consistent enrichment | implicit in coverage + audited via shadow comparison |
| Stable schemas | schema drift + configuration rollout status |
The serving SLO axis (latency + error rates) sits alongside as the tenant-facing layer — distinct from quality but on the same dashboard set because both shape consumer experience.
Why multi-dimensional matters¶
A single "data quality" metric doesn't surface the trade-offs. A high-freshness / low-completeness sequence is right for online inference but wrong for training. A schema-stable / enrichment-inconsistent sequence is right for backwards compatibility but introduces silent online-offline discrepancy. Treating quality as a vector forces the platform team and consumers to choose explicitly which dimensions matter for which workload, and lets the platform expose distinct freshness / completeness profiles per workload.
Seen in¶
- sources/2026-05-21-pinterest-making-user-sequence-data-more-cost-efficient-faster-and-easier-to-use — first canonical four-dimension framing for user event sequences in a multi-tenant ML platform.