New in 0.12
DRE 0.12 tightens configuration and makes automation contracts explicit. It remains pre-1.0; read the stability contract for the scope of the future 1.0 promise.
Configuration errors surface before delivery
Section titled “Configuration errors surface before delivery”Unknown keys, ambiguous selectors and unknown --set names fail rather than silently changing
the selection or ignoring configuration. Prefer explicit tag:<name> and name:<name>
selectors when a bare name could mean several things. Declare a query’s SQL in report YAML;
use project allow_unmanaged: true for SQL intentionally outside that managed list. Invalid DRE_RUN_DATE
values fail startup instead of falling back to today’s date. See Building reports
and environment variables.
dre init stores secret environment-variable references rather than secret values. It refuses
to replace unreadable or invalid profiles. Continue using env_var() and the runner’s secret
store for passwords, tokens and keys.
Data checks and schedules
Section titled “Data checks and schedules”Data tests arrived in 0.11. dre test executes fresh report queries and attached data_tests
without report formatting or delivery. dre build combines checks and delivery; error-severity
failures block the affected Binding. dre run deliberately skips those checks.
In 0.12, scheduled occurrence commands use dre build, so attached checks run before
delivery. dre run --schedule and dre build --schedule refuse paused schedules, and scheduled
instants must match the schedule’s firings. Select reports directly for an intentional manual
run outside that schedule. See Build and run reports
and Schedules.
Machine-readable contracts
Section titled “Machine-readable contracts”Consumers must recognize the current schema identifiers:
| Document | Schema |
|---|---|
| Manifest | dre/manifest/v4 |
| Run results | dre/run-results/v2 |
| JSON log events | dre/events/v1 |
| Validation JSON | dre/validate/v1 |
Binding lifecycle statuses are succeeded, failed, compiled, checked, cancelled and
timed_out; delivery states and data-test verdicts have their own vocabularies. Manifest
outputs[] owns output destinations; the old first-output aliases are gone. Compiled plans
also omit their old root format/output/destination aliases. A busy selected Binding exits
75, so orchestrators can distinguish it from ordinary failure. Update consumers against
the machine-output reference, logging and exit codes.
--output-name refuses a per-row first output instead of silently ignoring the override.
Use the output’s per-row naming configuration instead.
Plugin-owned format and SQL behavior
Section titled “Plugin-owned format and SQL behavior”Core no longer infers file extensions, query grouping, workbook rules or document context
from a format’s name. File plugins must advertise format_properties. The minimum versions
are CSV 1.2.0, document 1.6.0, fixed_width 1.1.0, parquet 1.1.0 and XLSX 1.2.0; install them
before upgrading core.
Read-only query validation uses the source’s sql_dialect policy. DuckDB 1.4.0, PostgreSQL
1.6.0 and Databricks 1.4.0 supply that policy and native catalogue discovery. A source without
it cannot serve dre query, even when it supports a read-only connection; this currently
includes ClickHouse and generic JDBC. JDBC’s native catalogue discovery remains available.
No missing policy falls back to generic SQL validation. See the plugin protocol
for custom-plugin migration and the exact metadata contract.
Legacy configuration and storage
Section titled “Legacy configuration and storage”The sources: alias for profile connections:, packages.yml as a dependency filename,
misplaced plugin declarations and old plugin install conventions are removed. Protocol peers
must support protocol 1. Old run files remain untouched rather than being migrated
automatically. Follow the migration inventory for every replacement and lock
checksum conversion; do not guess an old lock entry’s platform.
Target folders need reliable renames; atomic local delivery with if_exists: error or number
also needs hard links. secret_columns masks diagnostics and saved result copies, including
Arrow snapshots; it does not change delivered files. Temporary query spools contain original
values while the run executes. Restrict target-path access and retention, or disable per-row
saved results where appropriate. See The target path and
Password-protected outputs.