Hardened by real data
PGStream was tested against a real 83-table production backup, and every failure became a feature: zero dates, invalid enum markers, UUID-shaped keys that are not UUIDs, orphaned foreign-key rows, NUL bytes, and unsigned bigint keys are all detected in the data and handled with explicit, lossless-first rules.
Fail closed, hand over the fix
Unsupported conversions stop the run rather than silently changing the schema, and every stop ships its own remediation — the query to find the bad rows, the exact ALTER to reconcile a drifted target, or the cast flag that opts into a different mapping. The dry run surfaces all of it before a single write.
Measured, not assumed
Every completed table logs a time breakdown showing whether source reads, target writes, or transformation dominated. Wire compression and read/write pipelining took a two-million-row table from 23 minutes to under 10 over the same link, five parallel workers now move the complete 8.5-million-row database end to end in about the same time, and the published benchmarks include the configurations that did not help.