A green pipeline is worthless if one value changed meaning.
A data pipeline is only trustworthy when full-load values, counts, and recovery are explicit and every CDC reader declares what it emits. Omni Loader keeps uncertainty visible and gives each movement path a contract you can test against the records you will actually receive.
trialstarter{...}activescaleUNCHANGED_TOASTBurst Load proves the complete dataset contract
Counts, mapped values, precision, timestamps, LOBs, target DDL, and representative queries belong in the acceptance plan. Resumable slices retain explicit progress, and partial target work is cleaned before retry instead of being counted twice.
Unknown never becomes NULL by accident
PostgreSQL marks unchanged TOAST values explicitly. SQL Server uses typed references for unresolved off-row LOB, page-dictionary, and encrypted values. Oracle DirectRedo emits typed scalar values while unresolved LOB and object values remain bare-string markers. The representation changes by reader, but the uncertainty stays visible.
Before-images come with the prerequisite attached
MySQL and both Sybase readers emit before- and after-images. Oracle’s readers need supplemental logging ALL for complete old values; DirectRedo reconstructs update before-images from the logged undo. PostgreSQL depends on replica identity. Every Omni Loader SQL Server mode is after-image-only.
Schema behavior is declared per reader
PostgreSQL emits relation-diff schema events and first-class TRUNCATE. MySQL, Oracle LogMiner, and SQL Server surface their schema changes through their reader contracts. The raw Sybase strategies do not claim unified schema events, so this page does not imply that they do.
Initial load and CDC share a deliberate handoff
A bounded worker pool plans and copies table slices in parallel. PostgreSQL’s passthrough path holds one exported snapshot across its slices. Materialized paths use replay-safe overlap and idempotent target behavior because each engine offers a different snapshot primitive.
Durability has an address
Committed changes enter an agent-local durable spool with source positions attached. Independent sink cursors advance from that spool, and pause/resume keeps the checkpoint. A destination only moves its own cursor after its apply contract is satisfied.
The contract belongs to the exact reader
DirectRedo, SQL Server Streaming, and DirectDevice do not borrow correctness claims from their connection-based alternatives. Every capture path documents its own prerequisites, unresolved-value behavior, schema scope, and operating limits.
Proof follows the reader
Ten capture paths. Ten explicit contracts.
Oracle DirectRedo · Oracle LogMiner · SQL Server Streaming · SQL Server native CDC · SQL Server polling · Azure SQL native CDC · PostgreSQL pgoutput · MySQL binlog · Sybase DirectDevice · Sybase reconstruction
These ten paths do not share one blanket record promise. Log access, replica identity, supplemental logging, CDC objects, and storage layout can change what a reader can emit. The capture-mode page names those conditions before they become migration surprises.
One deliberate boundary deserves extra clarity: Sybase’s physical row identity can move under page compaction. Its proofs target no-loss behavior in those edge cases rather than claiming exact per-record identity under every storage rewrite.
Put your hardest row in the test plan.
Bring the LOBs, replica-identity settings, supplemental-logging posture, schema changes, and transaction patterns you cannot afford to misread. We will map each one to the exact reader contract and test the cases that carry the most risk.