[BETA] feat(nodespec): the single unified dataflow spec (replaces standard/flow/materialized_view)#104
Open
haillew wants to merge 17 commits into
Open
[BETA] feat(nodespec): the single unified dataflow spec (replaces standard/flow/materialized_view)#104haillew wants to merge 17 commits into
haillew wants to merge 17 commits into
Conversation
Introduce the nodester dataflow spec: a single, node-graph format
(source -> transformation -> target) that replaces the separate standard,
flow, and materialized-view formats. The transformer lowers a node graph
into the framework's existing flow spec, so all current capabilities (CDC,
snapshots, data quality, quarantine, sinks, table migration, materialized
views) are preserved.
Highlights:
- Target nodes wire inputs via an `input` list; each item is a node name
(auto flow name) or `{view, flow}` to define the SDP flow name explicitly
and keep it stable across edits (renaming a flow forces a full refresh).
- A single spec may contain both streaming-table and materialized-view
targets, including chains where one feeds the other.
- Python transformation nodes become their own view that applies
apply_transform to their inferred upstream; inline python_transform on a
source is still supported for backward compatibility.
- Inline SQL/Python sources (and append_sql) remain supported but warn, in
both the nodester and legacy formats; the recommended alternative is a
dedicated transformation node.
- Materialized view targets no longer accept an inline source_view (breaking);
chain a source node via `input` instead.
- All 38 nodester samples updated and verified end to end on Databricks.
- Adds ADR-0007 and a rewritten nodester spec reference.
Co-authored-by: Isaac
The local .claude directory (Claude Code commands/settings) should not be part of the repository. Add .claude/ to .gitignore and remove the previously tracked command files from version control. The files remain on disk locally (now ignored), so they persist across branch switches. Co-authored-by: Isaac
Move the feature-test GitHub workflow and the pattern-samples validation notebooks (validate_run_1..4 + validation_utils) out of this branch. They are kept on a local-only branch and intentionally not published upstream. Co-authored-by: Isaac
Rename the spec type, schema (spec_nodespec.json + nodespecSpec), transformer (NodespecSpecTransformer), migration script, sample bundle (nodespec_sample), pipelines, docs, and the data_flow_type value from "nodester" to "nodespec". No behavioural change — purely a rename. Co-authored-by: Isaac
haillew
changed the title
1 task
Rewrite the nodespec transformer around a single snake_case -> camelCase key map plus flat/recursive converters, replacing the per-context allowlist maps and per-source-type branches. Builders now copy every non-structural config key as passthrough detail. Output is unchanged (verified identical across all 38 samples and end to end on Databricks); ~593 -> 440 lines. Co-authored-by: Isaac
nodespec is the framework's unified spec format, so specs may now omit data_flow_type entirely. - DataflowSpecBuilder defaults data_flow_type to "nodespec" at read time, so type selection, metadata validation, schema validation, and transformation all treat a type-less spec as nodespec. - main.json routes any spec without an explicit legacy dataFlowType (standard/flow/materialized_view) to the nodespec schema. - spec_nodespec.json no longer requires data_flow_type; the const remains so that, when present, it must be "nodespec". Co-authored-by: Isaac
… renumber ADR to 0008 Field syntax (discriminated config for IntelliSense + consistent nesting): - Targets wire inputs via `input_flows`. - Data quality is a nested `data_quality` object; quarantine is a nested `quarantine` object (mode + optional target). Sink targets stay flat. - Sources allow an optional inline `python_transform` (transformer already supports it). Transformer reads the new field names. migrate_to_nodespec.py emits the new shape, fixes historical-snapshot conversion (no spurious source node) and table-migration source_details casing, and writes configs in a canonical field order (idempotent normalise for existing nodespec specs). Samples reordered to the canonical field order. Reference doc, feature-changes doc updated. Renumber ADR 0007 -> 0008 (0007 is taken by scripted-versioned-docs on main), with an amendment mapping the original syntax to the shipped names. Co-authored-by: Isaac
docs/nodespec_feature_samples_changes.md was an internal implementation/notes doc that shouldn't live in the public repo. The authoritative user-facing reference remains docs/source/dataflow_spec_ref_main_nodespec.rst and the decision record docs/decisions/0008-unified-nodespec-dataflow-spec.md. Co-authored-by: Isaac
The nodespec schema's $defs identifiers (sourceDeltaConfig, dataQuality, inputFlows, …) were camelCase while everything a nodespec spec author writes is snake_case. These names are schema-internal ref targets — they never appear in an authored spec — but renaming them to snake_case makes the nodespec schema fully self-consistent. - Renamed all nodespec $defs keys and their internal #/$defs refs to snake_case. - Updated main.json's external ref (#/$defs/nodespecSpec -> nodespec_spec). - The external ref out to definitions_main.json#/definitions/cdcSettings is unchanged (that def's name is definitions_main's concern). Legacy standardSpec/ flowsSpec/materializedViewsSpec refs are unaffected. No behavioural change: authored specs and the transformer are untouched; 37/37 nodespec samples still validate. Co-authored-by: Isaac
recursiveFileLookup was the last camelCase field an author writes in a nodespec spec. Renamed to recursive_file_lookup in the schema and samples; the transformer maps it back to the backend's `recursiveFileLookup` at build time (added to the snake->camel key map), and the conversion script snakes it during migration. Remaining camelCase in the schema is deliberately unchanged: JSON Schema keywords (additionalProperties, allOf, ...), literal Spark/Auto Loader option keys inside reader_options, framework-wide identifier values (source_type cloudFiles/ batchFiles/deltaJoin, disableOperationalMetadata), and the external cdcSettings ref. Co-authored-by: Isaac
Authors now write snake_case for nodespec's own enum values: - source_type: cloudFiles -> cloud_files, batchFiles -> batch_files, deltaJoin -> delta_join (delta/kafka/python/sql unchanged). - config_flags: disableOperationalMetadata -> disable_operational_metadata. Nodespec-only: the transformer maps these back to the backend's expected values (cloudFiles / disableOperationalMetadata) on output, so the backend and the legacy standard/flow/materialized_view formats are unchanged. The conversion script snakes them during migration. Left as-is (not nodespec-owned): reader_options keys AND values are literal Spark/Auto Loader tokens (e.g. binaryFile, addNewColumns, cloudFiles.format); the engine reads them verbatim. Co-authored-by: Isaac
Matches the schema/samples: source_type values are now cloud_files / batch_files / delta_join (delta/kafka/sql/python unchanged). Co-authored-by: Isaac
…_type Within target_delta_config, an intra-object if/then/else keyed on table_type now enforces which fields are valid: - table_type=mv: sql_path/sql_statement/refresh_policy/table_details allowed; cdc_settings/cdc_snapshot_settings/table_migration_details/once rejected. - table_type=st (or omitted): the reverse. table_type and the gated fields living in the same object is fine — if/then conditions on one property and constrains its siblings. Editors now flag, e.g., sql_path on a streaming table or cdc_settings on a materialized view. 37/37 samples still validate. Co-authored-by: Isaac
Extend the if/then pattern to other genuine field dependencies: - source_delta_config: starting_version_from_dlt_setup=true requires cdf_enabled=true. - target_delta_config: cluster_by_columns and partition_columns are mutually exclusive. - snapshot_cdc_config: periodic snapshots forbid inline source_type/source (they read an upstream source node); historical snapshots require them; a file source requires path+format; a table source requires table. Sinks are intentionally left flat/ungated for now (the sink redesign is deferred, and target_type lives at the node level rather than in the sink config). 37/37 samples still validate. Co-authored-by: Isaac
…n exclusivity Reflects the schema's conditional validation: streaming-table-only vs materialized-view-only target fields are gated on table_type, and partition_columns / cluster_by_columns are mutually exclusive. Co-authored-by: Isaac
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What this PR adds (BETA)
A single, unified, node-based dataflow spec — nodespec — that supersedes the standard, flow, and materialized-view formats.
A nodespec is a graph of nodes that chain together:
Targets declare what feeds them via an explicit
inputlist; sources and transformations reference their upstream within their own definition. The node graph lowers into the framework's existing flow-spec representation, so no engine changes are required and every current capability (CDC, snapshots, data quality, quarantine, sinks, table migration, materialized views) is preserved. A single nodespec can mix streaming-table and materialized-view targets, including chains where one feeds the other.Legacy formats continue to work, and
scripts/migrate_to_nodespec.pyconverts existing standard/flow/materialized_view specs.Why
The three separate formats force users to learn three layouts, leak engine internals (flow groups, view registration, staging tables), nest settings deeply, hide pipeline topology, and force streaming tables and materialized views into separate files. nodespec collapses all of that into one readable, chainable model — easier for newcomers writing their first pipeline and for large teams reading and editing each other's pipelines.
Details
See the design decision record:
docs/decisions/0008-unified-nodespec-dataflow-spec.md(full rationale, before/after examples, and the key decisions).