mdcode: binding profiles for the semantic model - #358
Merged
libei merged 8 commits intoAug 30, 2026
Conversation
…ader The binding-profiles guide authors models with a readable surface the loader did not accept. Teach the loader that surface so a single combined file using it loads correctly (profiles themselves come next): - entities: is an alias for datasets: - a model-level deployment_target: URI folds into the GOOGLE custom_extensions block the deploy leg reads (the existing form still works; the two must agree when both are present) - a field expression: may be a one-line string, shorthand for a single target-dialect variant - a field may be marked unbound: true -- declared logically but with no physical column under this binding, structurally absent rather than null Sugars are normalized before schema validation; unbound is carried onto the IR. A field that is neither bound nor unbound is a load error naming the field.
Adds the mechanism behind binding profiles: one logical model paired at push time with a chosen physical binding, deployed to BigQuery + Knowledge Catalog as before. Deploy legs are unchanged; this is loader/push work. - Profile discovery: SemanticModelLayout reads <model>.profiles/*.yaml beside the model (single-level, so a profile file is never taken for a model). - Merge (resolve_profiles.mergeProfile): overlays a profile onto the logical model by name -- entity source, field expression, deployment target -- and enforces the binding-only contract (a profile may set physical bindings and may leave a field unbound; it may not add/remove elements or change what anything means). A field the profile omits is carried through as unbound. - Availability pruning (resolve_profiles.pruneUnavailable): drops each unbound field and everything that depends on it -- a metric that reads it, a relationship whose join column is unbound, a cross-entity metric over a dropped relationship -- and returns a per-profile report of withheld coverage. - kcmd push --profile <name>: merges the named profile before load, prunes after, then runs the existing validation + deploy. Orthogonal to --target. The implicit 'default' profile keeps today's inline single-file behavior. - Sources may be resource-name URIs: a //bigquery.googleapis.com/... table is normalized to project.dataset.table; a non-BigQuery URI (Spanner/AlloyDB/...) rides through and is skipped by the BigQuery pre-flight probe. Tested: mergeProfile (by-name merge, contract violations, omission->unbound, non-mutation), pruneUnavailable (field/metric/relationship pruning + report), profile discovery, and an end-to-end run of the guide's three-file example.
Completes binding profiles with the reporting surface and the CI default. - kcmd profiles: lists each of a model's binding profiles and, per profile, its resolved deployment target and sources plus what it cannot answer (the unbound fields and the metrics/relationships that fall with them). Read-only -- it merges and prunes each profile the way push does but deploys nothing and makes no network call -- so coverage can be inspected before choosing a profile. - default_profile in catalog.yaml: the profile a bare kcmd push uses when --profile is omitted. Precedence is --profile, then default_profile, then the inline 'default' binding. Round-trips through manifest load/save. Tested: the profiles command's listing, default marking, resolved sources, and withheld coverage; default_profile load/save round-trip.
The BigQuery Graph generator used a field's logical name at every structural site -- node KEY, edge SOURCE KEY / DESTINATION KEY / REFERENCES, association REFERENCES -- and as the operand of a MEASURE. That was invisible while every binding named its column after the field (name == column, as all fixtures do), but a binding profile that maps a field to a differently named physical column made the generator emit the logical name where BigQuery requires the physical column, so the CREATE PROPERTY GRAPH failed with "Column not found" and "Unrecognized name". Resolve each structural column to its bound physical column (physicalColumn / physicalColumns), and let a MEASURE aggregate a declared field by its exposed property name even when the binding renamed the underlying column. Output is unchanged when name == column, so existing goldens do not move.
libei
changed the title
The BigQuery generator's physical-column resolution had no offline test:
every fixture bound its columns name == column, so reverting the fix left the
whole suite green while a renamed profile binding failed on the live backend
("Column not found", "Unrecognized name"). This adds the missing regression.
A binding profile is two files, so it does not fit the single-file corpus
golden loop. Add fixtures/profiles/profile_binding.yaml (logical declarations)
and profile_binding.analytical.yaml (a profile that renames every physical
column and leaves one field unbound), run the same merge -> prune -> IR -> DDL
path push takes, and check the emitted DDL plus the availability report against
a committed golden. The golden shows the physical column at each structural
site (node KEY, edge SOURCE KEY / DESTINATION KEY / REFERENCES) and the
exposed property alias as each MEASURE operand -- the shape live BigQuery
accepts -- and the pruned field and metric. Reverting the generator fix now
fails this test.
The profiles/ subtree is sugared authoring input, not standalone OSI (a
logical model declares fields with no expression; a profile carries
deployment_target/unbound sugars), so the OSI-schema guardrail skips it; the
loader, merge, and this golden validate it instead.
The profile golden suite had one case: a purely logical model bound entirely
by a profile. It did not exercise a model that already carries inline bindings
and is then overridden by a profile -- the combined single-file form, which is
where a model starts before any profile exists.
Add a corpus of logical/profile pairs, each with its own golden, spanning the
distinct cases:
- profile_binding + analytical : logical base, profile supplies all bindings
and renames every column (the existing case).
- profile_binding + operational: the same logical base bound a second way, a
different field unbound -- one model, two physical realizations.
- partial_binding + prod : a combined file (fully bound inline) whose
profile overrides only the sources and target and inherits every omitted
column (an environment swap).
- partial_binding + remap : the same combined base, a profile that
renames one column and unbinds another, inheriting the rest.
The prod golden pins the inherit-on-omission rule: a field a profile does not
mention keeps its base binding rather than becoming unbound.
One canonical predicate for "is a field bound" (ir.ts fieldBinding), used by both availability pruning and the BigQuery generator so they never disagree: - A field awaiting transpilation (column carried on importedExpression, target expression not yet filled) is bound, not unbound. Prune no longer drops such fields or the metrics over them. - An unbound KEY field now makes the whole entity unavailable: prune drops the entity and cascades to its relationships and metrics, and reports it (AvailabilityReport.droppedEntities), instead of emitting an unkeyed node that only fails at BigQuery deploy. - A structural site (KEY / SOURCE KEY / DESTINATION KEY / REFERENCES) bound to a computed expression now warns statically rather than emitting invalid DDL. - push and profiles share one mergeProfileOntoDoc helper; profiles now surfaces merge warnings (it silently dropped them before). Adds regression tests for each: imported-only field stays bound, unbound key drops the entity and its dependents, and a computed key warns.
Drop the proposed/not-implemented banner for a scope note; the feature ships (loader sugars, merge, prune, --profile, kcmd profiles, default_profile) with BigQuery Graph + Knowledge Catalog as the only deploy targets. Correct claims that overran the code: - profiles merge only entities and fields by name; relationships and metrics never appear in a profile (mergeProfile rejects them). - a profile cannot rebind a relationship's junction source. - availability prunes entities (unbound key drops the whole node), metrics, and relationships; there is no action pruning. - validation probes the source table, not columns; a mistyped column surfaces at deploy, not at validate. - push prints a per-profile availability summary; kcmd profiles lists each withheld item. - a Spanner-bound profile is merged and reported but not deployed.
libei
added a commit
that referenced
this pull request
Aug 30, 2026
…ath segment (#366) The Spanner graph generator reduced a source to its final DOTTED segment, which is correct for a `project.dataset.table` reference but mangles a resource-name URI: a binding profile's `//spanner.googleapis.com/.../tables/Customer` source (kept verbatim by the loader, unlike a BigQuery URI which is rewritten to dotted form) emitted `com/projects/.../tables/Customer` as the table name, which Spanner rejects. Reduce a resource-name (or any scheme://) URI to its final PATH segment first, mirroring the loader's BigQuery-URI handling. Localized to the Spanner leg so the Knowledge Catalog leg still records the full resource name. Surfaces alongside binding profiles (#358) + Spanner push (#361): profiles.md's operational example uses spanner:// sources and would fail live.
libei
added a commit
to libei/knowledge-catalog
that referenced
this pull request
Aug 30, 2026
Rework codelab section 5 and update the binding-profiles guide to reflect the Spanner Graph target (GoogleCloudPlatform#361) and binding profiles (GoogleCloudPlatform#358) together. Codelab: - Step 1 now names the deployment target with the first-class `deployment_target` key instead of a GOOGLE custom_extensions JSON string. - Section 5 replaces the `sed` target-swap with a real Spanner binding profile (`sales.profiles/operational.yaml`): different table names, remapped columns (including keys), and an unbound `net_amount` so `revenue` is withheld operationally. It shows `kcmd profiles`' availability report, the generated Spanner DDL, a live deploy, and a GQL query -- every output captured from a live ENTERPRISE Spanner run. Cleanup drops the Spanner database. profiles.md: - Scope note updated: Spanner graphs now deploy (was "non-BigQuery not yet supported"). - Spanner sources use bare table names, not `//spanner.../tables/X` resource URIs, which a Spanner graph cannot reference; added a note on the rule. - `--target` gains `spanner`; command examples deploy the operational (Spanner) profile.
libei
added a commit
that referenced
this pull request
Aug 30, 2026
* docs(mdcode): codelab deploys to Spanner via a binding profile Rework codelab section 5 and update the binding-profiles guide to reflect the Spanner Graph target (#361) and binding profiles (#358) together. Codelab: - Step 1 now names the deployment target with the first-class `deployment_target` key instead of a GOOGLE custom_extensions JSON string. - Section 5 replaces the `sed` target-swap with a real Spanner binding profile (`sales.profiles/operational.yaml`): different table names, remapped columns (including keys), and an unbound `net_amount` so `revenue` is withheld operationally. It shows `kcmd profiles`' availability report, the generated Spanner DDL, a live deploy, and a GQL query -- every output captured from a live ENTERPRISE Spanner run. Cleanup drops the Spanner database. profiles.md: - Scope note updated: Spanner graphs now deploy (was "non-BigQuery not yet supported"). - Spanner sources use bare table names, not `//spanner.../tables/X` resource URIs, which a Spanner graph cannot reference; added a note on the rule. - `--target` gains `spanner`; command examples deploy the operational (Spanner) profile. * docs(mdcode): make the codelab profile-first from step 1 Author the model as a logical model plus a named `analytical` binding profile from the start, instead of an inline single-file binding, so the profile feature is the default path a reader learns. Set it as `default_profile` so steps 3-4 keep using a bare `kcmd push`; section 5 then adds the `operational` Spanner profile as a second binding rather than introducing the concept. Add a "simple case" callout showing the single-file inline binding (the `default` profile) for models that only ever bind to one store. Update the section-5 `kcmd profiles` output to show both profiles side by side (analytical withholds nothing; operational withholds revenue). Verified live: `kcmd profiles`, and `kcmd push --target kc|bq --validate-only` off the split layout produce the documented output. * docs(mdcode): restructure codelab as author-logical -> govern -> bind Reorder the codelab to match how a semantic model is actually adopted, now that a logical-only model can be pushed to Knowledge Catalog (#367): 1. Author the logical model (no source, no column, no target) 2. Govern it in Knowledge Catalog -- pure logical, needs no tables or data 3. Deploy to BigQuery -- add the analytical binding, hydrate, query 4. Deploy the same model to Spanner -- add the operational binding The Knowledge Catalog push moves ahead of hydration and binding: it governs meaning, so it needs neither. The analytical binding authoring and the table hydration fold into the BigQuery step, where they are actually required. The "why tables come first" note is rescoped to the BigQuery push, which does validate source tables, unlike the KC push. Deploy prose drops "...Graph" (deploy to BigQuery / to Spanner); the graph names stay in the generated DDL, command output, and feature references. The Spanner leg now states explicitly that Knowledge Catalog is not re-pushed -- a binding profile changes nothing logical. * docs(mdcode): drop profiles.md changes from this branch (handled separately) * docs(mdcode): soften KC framing so it doesn't preclude governing bindings * docs(mdcode): state the develop/govern/bind arc plainly, drop defensive framing * docs(mdcode): drop redundant no-bindings-yet paragraph in step 1 * docs(mdcode): fix intro claim that the metric computes identically everywhere * docs(mdcode): intro no longer implies KC governs only the logical model * docs(mdcode): use entities: throughout and make outputs match live runs Verified the codelab end to end against live instances (Knowledge Catalog on autopush, BigQuery, and an ENTERPRISE Spanner instance) and reconciled it with what the commands actually print. - Author the model and both binding profiles under `entities:` rather than `datasets:`, matching the prose, which calls them entities throughout. The two keys are aliases; the OWL-import output still shows `datasets:` (the original spelling) and a note points that out. - Step 2's Knowledge Catalog plan now shows the real output: the `Validating...` framing, the two relationship-name-normalization warnings (with a note that they are expected and repeat on every KC push), and the completion footer. - Step 3's BigQuery DDL block shows the `Pushing...` line, the `-- BigQuery Graph --` banner, and the deployment-target comment it was missing, matching the Spanner block in step 4 and the real console output. Depends on the logical-only KC emitter fix (source.resources) so step 2's govern-before-you-bind push succeeds against the live server.
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.
Adds binding profiles to the kcmd semantic model: one logical model
(declarations) paired at push time with a chosen physical binding profile
via
kcmd push --profile <name>. This relaxes today's 1:1 entity→sourcebinding so a single model can deploy against different stores or physical
layouts, and lets a profile bind only the subset of fields its store serves.
The single-file (inline
default) form keeps working unchanged. Deploy legsare unchanged — BigQuery Graph + Knowledge Catalog only.
What's here (four commits)
entities:alias for
datasets:, first-classdeployment_target:, bare-stringexpression:, andunbound: trueon a field. Sugars normalize pre-Zod, sothe IR and validation stay untouched.
<model>.profiles/*.yamldiscovery;mergeProfileoverlays a profile's physical bindings onto the logical model by name (with a
binding-only guard);
pruneUnavailabledrops fields a profile leaves unboundand propagates that up to the metrics/relationships built on them, with a
per-profile report.
kcmd push --profile <name>is orthogonal to--target.kcmd profiles(read-only listing: target, sources,"cannot answer" report) and
default_profile:in catalog.yaml.at structural sites (node KEY, edge SOURCE/DESTINATION KEY, REFERENCES) and
as MEASURE operands. That was invisible while name == column but broke a
profile that renames a column. Now resolved to the bound physical column;
output is unchanged when name == column, so goldens do not move.
Validation
bun testundertoolbox/mdcode: full semantic + tool suites green, nogolden churn.
BigQuery Graph.
analytical(all fields bound) answersorder_countandtotal_items;lite(item_countunbound) answersorder_countandcorrectly rejects
total_itemsas pruned. BothCREATE PROPERTY GRAPHstatements succeed and the measures return through
GRAPH_EXPAND+AGG.Supersedes the stacked split; #359 and #360 fold in here.