Skip to content

CONCEPT Cited by 1 source

CDC bad-data propagation

Definition

CDC bad-data propagation is the structural hazard of any change-data-capture pipeline: because the target-table state at time T+1 is computed from the target-table state at time T plus the deltas applied between them, a corrupted target-table partition at time T becomes embedded in every subsequent state without separate intervention.

"Being a CDC process means the data generated by the system is used again to generate the new data. This means if previous landed data has any issues the problematic data will be passed to the new landed data." — Source: sources/2026-05-12-meta-migrating-data-ingestion-systems-at-meta-scale

Why it matters

This property is structurally different from non-CDC migrations:

  • In batch-snapshot systems, a single bad batch can be discarded and the source re-read; the canonical state is always reachable from the source.
  • In CDC systems, the canonical state is incrementally reconstructed from a base snapshot plus applied deltas. A corrupted delta or target partition becomes part of the computation that produces the next state.

Consequence: any data-quality incident in a CDC pipeline has a propagation horizon — the time window between corruption and detection bounds the rollback work required, because every target partition produced after the bad delta carries the corruption forward.

Containment primitives

Two operational primitives bound the blast radius of a CDC data-quality incident:

  1. Detection-side: continuous quality checks against parallel reference data (e.g. a previous-system shadow output, or a periodic full re-snapshot from source) so corruption is caught within bounded latency rather than discovered by downstream consumers later.
  2. Containment-side: partition-level quality marking — annotate the corrupted partition's metadata with a bad-quality flag; a delta partition so marked stops new data landing + alerts an operator; a target partition so marked is substituted with an older known-good partition merged forward with additional deltas, bypassing the corrupted slice without consumer impact.

CDC-migration framing

CDC-system-to-CDC-system migration doubles the propagation surface: two pipelines run side by side, each subject to its own bad-data-propagation risk. Meta's patterns/shadow-then-reverse-shadow-migration shape exists in part because both pipelines need ongoing data-quality signal: during the reverse shadow phase the original production job, now demoted to shadow-writer, is still around to provide the comparison reference if the new production-writer corrupts a partition.

  • vs concepts/blast-radius at the partition layer: bad-data propagation is the temporal shape of blast radius — corruption spreads forward through time, not just outward through systems.
  • vs concepts/eventual-consistency: eventual consistency is about temporary divergence from a converging canonical state; bad-data propagation is about permanent divergence unless the corruption is explicitly remediated.
  • vs concepts/replication-lag: replication lag is the delay before correct data lands; bad-data propagation is about incorrect data landing and being used as the basis for subsequent computation.

Seen in

Last updated · 542 distilled / 1,571 read