Instrument package-manager detection telemetry at project setup

Why:
We need first-party data on which Python package manager(s) our users'
projects actually use (pip/conda/uv/poetry) to prioritize VPEX setup-flow
investment, replacing public-survey estimates. Measurement only -- no setup
behavior changes.

What:
- Add packageManagerDetection.ts: a pure, signal-based classifier that
  reports all applicable managers plus a best-guess primary (uv > poetry >
  conda > pip), the firing signals, hasLockfile, and interpreter source.
  Treats bare uv/poetry on PATH as weak signals.
- Add Events.PYTHON_ENV_SETUP_DETECTED with a typed, documented schema in
  telemetry/constants.ts (reuses existing Telemetry client; opt-out honored;
  categorical data only, no paths/package/cluster names).
- Add telemetry/packageManagerExtensions.ts: the emit half, layered onto the
  Telemetry class via the commandExtensions declare-module pattern
  (recordPackageManagerDetection). Keeps disk/Python-extension deps out of the
  Telemetry client.
- Add PackageManagerTelemetry.ts: the collection half -- a best-effort,
  non-blocking collector (disk + already-resolved interpreter metadata) that
  gathers signals, runs the pure classifier, and calls the emit method.
  Deduplicated per session on (trigger, projectRoot); failures degrade to
  unknown and are swallowed.
- Wire emission into three touchpoints: project-open env check (auto_open),
  the set-up-environment command (explicit_command), and first Run/Debug
  with Databricks Connect (run/debug).
- Add unit tests for the detector and pure helpers, and a dashboard-owner
  handoff note.

Detection correctness:
- interpreterSource is derived from the active interpreter alone, never from
  project files: a uv.lock project on a conda/venv/system interpreter reports
  that interpreter's real source, keeping the setup-flow gap visible. A
  genuinely uv-provisioned venv is identified by the `uv =` marker in
  pyvenv.cfg (pure pyvenvCfgMarksUv), not by uv.lock.
- conda is attributed only when the active interpreter resides under
  CONDA_PREFIX (pure interpreterUnderCondaPrefix, with a path-boundary check),
  not on the bare env var, which is session-global in the extension host
  (launching from an activated conda shell) and would otherwise over-count
  conda for uv/poetry/pip projects.
- pyproject [tool.uv]/[tool.poetry] detection uses a pure, bounded table-header
  scan (pyprojectHasToolSection) instead of substring matching: ignores
  comments and in-value mentions, rejects prefix collisions (e.g. tool.uvicorn),
  and matches subtable and array-of-table headers (e.g. [tool.uv.sources],
  [[tool.poetry.source]]) that the substring check missed.
- No external executable is run for telemetry: the uv-on-PATH probe was
  removed (it spawned a PATH-resolved `uv` for a weak, non-attributing signal);
  detection now only reads disk and already-resolved interpreter metadata.

Verification:
- yarn run build (typecheck) passes.
- eslint clean; prettier formatted.
- yarn run test:unit: 202 passing, 0 failing (includes detector + helper tests).

Co-authored-by: Isaac

Author: rugpanov

packages/databricks-vscode/src/extension.ts

@@ -75,6 +75,7 @@ import {BundleVariableTreeDataProvider} from "./ui/bundle-variables/BundleVariab
 import {ConfigurationTreeViewManager} from "./ui/configuration-view/ConfigurationTreeViewManager";
 import {getCLIDependenciesEnvVars} from "./utils/envVarGenerators";
 import {EnvironmentCommands} from "./language/EnvironmentCommands";
+import {PackageManagerTelemetry} from "./language/PackageManagerTelemetry";
 import {WorkspaceFolderManager} from "./vscode-objs/WorkspaceFolderManager";
 import {SyncCommands} from "./sync/SyncCommands";
 import {CodeSynchronizer} from "./sync";
@@ -335,6 +336,23 @@ export async function activate(
         customWhenContext,
         telemetry
     );
+    const packageManagerTelemetry = new PackageManagerTelemetry(
+        telemetry,
+        pythonExtensionWrapper,
+        () => {
+            try {
+                return workspaceFolderManager.activeProjectUri.fsPath;
+            } catch (e) {
+                return undefined;
+            }
+        },
+        () => {
+            if (connectionManager.serverless) {
+                return "serverless";
+            }
+            return connectionManager.cluster ? "cluster" : "none";
+        }
+    );
     context.subscriptions.push(
         bundleFileWatcher,
         bundleValidateModel,
@@ -598,13 +616,15 @@ export async function activate(
                 connectionManager,
                 pythonExtensionWrapper,
                 environmentDependenciesInstaller,
-                configureAutocomplete
+                configureAutocomplete,
+                packageManagerTelemetry
             )
     );
     const environmentCommands = new EnvironmentCommands(
         featureManager,
         pythonExtensionWrapper,
-        environmentDependenciesInstaller
+        environmentDependenciesInstaller,
+        packageManagerTelemetry
     );
     context.subscriptions.push(
         telemetry.registerCommand(
@@ -982,7 +1002,8 @@ export async function activate(
         featureManager,
         context,
         customWhenContext,
-        telemetry
+        telemetry,
+        packageManagerTelemetry
     );
     const debugFactory = new DatabricksDebugAdapterFactory(
         connectionManager,

packages/databricks-vscode/src/language/EnvironmentCommands.ts

@@ -5,16 +5,19 @@ import {Cluster} from "../sdk-extensions";
 import {EnvironmentDependenciesInstaller} from "./EnvironmentDependenciesInstaller";
 import {Environment} from "./MsPythonExtensionApi";
 import {environmentName} from "../utils/environmentUtils";
+import {PackageManagerTelemetry} from "./PackageManagerTelemetry";
 
 export class EnvironmentCommands {
     constructor(
         private featureManager: FeatureManager,
         private pythonExtension: MsPythonExtensionWrapper,
-        private installer: EnvironmentDependenciesInstaller
+        private installer: EnvironmentDependenciesInstaller,
+        private packageManagerTelemetry: PackageManagerTelemetry
     ) {}
 
     async setup(stepId?: string) {
         commands.executeCommand("configurationView.focus");
+        void this.packageManagerTelemetry.emitDetection("explicit_command");
         await window.withProgress(
             {location: {viewId: "configurationView"}},
             () => this._setup(stepId)

packages/databricks-vscode/src/language/EnvironmentDependenciesVerifier.ts

@@ -10,6 +10,7 @@ import {ResolvedEnvironment} from "./MsPythonExtensionApi";
 import {NamedLogger} from "@databricks/sdk-experimental/dist/logging";
 import {ConfigureAutocomplete} from "./ConfigureAutocomplete";
 import {workspaceConfigs} from "../vscode-objs/WorkspaceConfigs";
+import {PackageManagerTelemetry} from "./PackageManagerTelemetry";
 
 export class EnvironmentDependenciesVerifier extends MultiStepAccessVerifier {
     private readonly logger = NamedLogger.getOrCreate(Loggers.Extension);
@@ -18,7 +19,8 @@ export class EnvironmentDependenciesVerifier extends MultiStepAccessVerifier {
         private readonly connectionManager: ConnectionManager,
         private readonly pythonExtension: MsPythonExtensionWrapper,
         private readonly installer: EnvironmentDependenciesInstaller,
-        private readonly configureAutocomplete: ConfigureAutocomplete
+        private readonly configureAutocomplete: ConfigureAutocomplete,
+        private readonly packageManagerTelemetry: PackageManagerTelemetry
     ) {
         super([
             "checkCluster",
@@ -403,6 +405,9 @@ export class EnvironmentDependenciesVerifier extends MultiStepAccessVerifier {
     }
 
     override async check() {
+        // First environment check on project open: emit package-manager
+        // detection telemetry (deduplicated per session, never throws).
+        void this.packageManagerTelemetry.emitDetection("auto_open");
         await this.connectionManager.waitForConnect();
         await Promise.all([
             this.checkCluster(this.connectionManager.cluster),

packages/databricks-vscode/src/language/PackageManagerTelemetry.ts

@@ -0,0 +1,251 @@
+import fs from "node:fs";
+import path from "node:path";
+import {NamedLogger} from "@databricks/sdk-experimental/dist/logging";
+import {Loggers} from "../logger";
+import {Telemetry} from "../telemetry";
+import {ComputeType, SetupTrigger} from "../telemetry/constants";
+import "../telemetry/packageManagerExtensions";
+import {MsPythonExtensionWrapper} from "./MsPythonExtensionWrapper";
+import {ResolvedEnvironment} from "./MsPythonExtensionApi";
+import {
+    detectPackageManagers,
+    interpreterUnderCondaPrefix,
+    InterpreterSource,
+    PackageManagerSignals,
+    pyprojectHasToolSection,
+    pyvenvCfgMarksUv,
+} from "./packageManagerDetection";
+
+export type {SetupTrigger};
+
+/**
+ * Collects package-manager signals at project-setup touchpoints and emits the
+ * {@link Events.PYTHON_ENV_SETUP_DETECTED} telemetry event.
+ *
+ * All probing is best-effort and non-blocking: any failure degrades to
+ * `unknown` and is swallowed, never thrown into the user's setup/run flow.
+ * Only categorical/enum data is emitted — no paths, package names, or other
+ * free-form content (see {@link detectPackageManagers}). Telemetry opt-out is
+ * honoured by the underlying {@link Telemetry} client.
+ */
+export class PackageManagerTelemetry {
+    private readonly logger = NamedLogger.getOrCreate(Loggers.Extension);
+
+    /**
+     * Triggers already emitted for the current `(project, trigger)` pair, to
+     * deduplicate within a session so one project open doesn't inflate counts.
+     */
+    private readonly emitted = new Set<string>();
+
+    constructor(
+        private readonly telemetry: Telemetry,
+        private readonly pythonExtension: MsPythonExtensionWrapper,
+        private readonly getProjectRoot: () => string | undefined,
+        private readonly getComputeType: () => ComputeType | "none"
+    ) {}
+
+    /**
+     * Detect the package manager(s) for the active project and emit telemetry.
+     * Deduplicated per `(project root, trigger)` within the session. Never
+     * throws.
+     */
+    async emitDetection(trigger: SetupTrigger): Promise<void> {
+        try {
+            const projectRoot = this.getProjectRoot();
+            if (projectRoot === undefined) {
+                return;
+            }
+            const dedupeKey = `${trigger}:${projectRoot}`;
+            if (this.emitted.has(dedupeKey)) {
+                return;
+            }
+            this.emitted.add(dedupeKey);
+
+            const env = await this.resolveEnvironment();
+            const signals = this.collectSignals(projectRoot, env);
+            const detection = detectPackageManagers(signals);
+
+            this.telemetry.recordPackageManagerDetection(detection, {
+                pythonVersion: this.getPythonMinorVersion(env),
+                targetCompute: this.getComputeType(),
+                trigger,
+            });
+        } catch (e) {
+            // Detection is measurement-only and must never disrupt setup.
+            this.logger.debug("Package manager detection failed", e);
+        }
+    }
+
+    private async resolveEnvironment(): Promise<
+        ResolvedEnvironment | undefined
+    > {
+        try {
+            return await this.pythonExtension.pythonEnvironment;
+        } catch (e) {
+            this.logger.debug("Failed to resolve python environment", e);
+            return undefined;
+        }
+    }
+
+    /** Detected interpreter minor version (e.g. "3.11"), if available. */
+    private getPythonMinorVersion(
+        env: ResolvedEnvironment | undefined
+    ): string | undefined {
+        const version = env?.version;
+        if (version?.major === undefined || version.minor === undefined) {
+            return undefined;
+        }
+        return `${version.major}.${version.minor}`;
+    }
+
+    /**
+     * Classify the *active interpreter's* provenance from the resolved
+     * environment alone. This is deliberately independent of project files: a
+     * project carrying `uv.lock` but running a conda/venv/system interpreter
+     * must report that interpreter's real source, so the setup-flow gap ("uv
+     * project, interpreter not uv-managed yet") stays visible. `uv.lock` is
+     * still captured as a strong *project* signal via `hasUvLock`.
+     */
+    private getInterpreterSource(
+        env: ResolvedEnvironment | undefined
+    ): InterpreterSource {
+        if (env?.environment === undefined) {
+            // No managed environment: a global/system interpreter.
+            return env ? "system" : "unknown";
+        }
+
+        const tools = env.tools ?? [];
+        if (env.environment.type === "Conda" || tools.includes("Conda")) {
+            return "conda";
+        }
+        if (
+            tools.includes("Venv") ||
+            tools.includes("VirtualEnv") ||
+            tools.includes("Poetry") ||
+            tools.includes("Pipenv") ||
+            env.environment.type === "VirtualEnvironment"
+        ) {
+            // The MS Python extension reports uv-created venvs as plain virtual
+            // environments. Distinguish a genuinely uv-provisioned interpreter
+            // by the `uv = <version>` line uv writes into pyvenv.cfg -- this is
+            // interpreter provenance, not the mere presence of uv.lock.
+            return this.isUvCreatedVenv(env) ? "uv" : "venv";
+        }
+        return "unknown";
+    }
+
+    /**
+     * True if the active venv's pyvenv.cfg marks it as uv-created. Thin fs
+     * wrapper around the pure {@link pyvenvCfgMarksUv}.
+     */
+    private isUvCreatedVenv(env: ResolvedEnvironment): boolean {
+        try {
+            const sysPrefix = env.executable.sysPrefix;
+            if (!sysPrefix) {
+                return false;
+            }
+            const cfg = path.join(sysPrefix, "pyvenv.cfg");
+            if (!fs.existsSync(cfg)) {
+                return false;
+            }
+            return pyvenvCfgMarksUv(fs.readFileSync(cfg, "utf-8"));
+        } catch (e) {
+            this.logger.debug("Failed to read pyvenv.cfg", e);
+            return false;
+        }
+    }
+
+    /**
+     * Gather raw signals from disk and the environment. Each probe is guarded
+     * so a single failure degrades that signal to absent rather than aborting.
+     */
+    private collectSignals(
+        projectRoot: string,
+        env: ResolvedEnvironment | undefined
+    ): PackageManagerSignals {
+        const exists = (file: string) => this.fileExists(projectRoot, file);
+        const interpreterSource = this.getInterpreterSource(env);
+        const pyproject = this.readPyproject(projectRoot);
+
+        const hasPyprojectToolUv = pyprojectHasToolSection(pyproject, "uv");
+        const hasPyprojectToolPoetry = pyprojectHasToolSection(
+            pyproject,
+            "poetry"
+        );
+        const hasPyprojectPipOnly =
+            pyproject !== undefined &&
+            !hasPyprojectToolUv &&
+            !hasPyprojectToolPoetry;
+
+        return {
+            hasUvLock: exists("uv.lock"),
+            hasPyprojectToolUv,
+            // uvOnPath is intentionally left unset: it is a weak signal that
+            // never attributes a project to uv, and probing it would mean
+            // executing a PATH-resolved `uv` binary purely for telemetry.
+            hasPoetryLock: exists("poetry.lock"),
+            hasPyprojectToolPoetry,
+            poetryOnPath: undefined,
+            hasRequirementsTxt: this.hasRequirementsTxt(projectRoot),
+            hasConstraintsTxt: exists("constraints.txt"),
+            hasPyprojectPipOnly,
+            hasCondaEnvFile:
+                exists("environment.yml") || exists("environment.yaml"),
+            hasCondaPrefix: this.hasActiveCondaInterpreter(env),
+            interpreterSource,
+        };
+    }
+
+    /**
+     * Whether the *active interpreter* lives under `CONDA_PREFIX`.
+     *
+     * We deliberately do NOT fire on the bare presence of `CONDA_PREFIX` /
+     * `CONDA_DEFAULT_ENV`: those are session-global in the extension host (set
+     * for every project when VS Code is launched from an activated conda
+     * shell), so using them directly would over-count conda for uv/poetry/pip
+     * projects. Requiring the active interpreter to reside under the prefix
+     * keeps this a project-scoped signal.
+     */
+    private hasActiveCondaInterpreter(
+        env: ResolvedEnvironment | undefined
+    ): boolean {
+        return interpreterUnderCondaPrefix(
+            env?.executable.sysPrefix,
+            process.env["CONDA_PREFIX"]
+        );
+    }
+
+    private fileExists(projectRoot: string, file: string): boolean {
+        try {
+            return fs.existsSync(path.join(projectRoot, file));
+        } catch (e) {
+            this.logger.debug(`Failed to stat ${file}`, e);
+            return false;
+        }
+    }
+
+    /** True if any `requirements*.txt` file exists in the project root. */
+    private hasRequirementsTxt(projectRoot: string): boolean {
+        try {
+            return fs
+                .readdirSync(projectRoot)
+                .some((name) => /^requirements.*\.txt$/.test(name));
+        } catch (e) {
+            this.logger.debug("Failed to list project root", e);
+            return false;
+        }
+    }
+
+    private readPyproject(projectRoot: string): string | undefined {
+        try {
+            const file = path.join(projectRoot, "pyproject.toml");
+            if (!fs.existsSync(file)) {
+                return undefined;
+            }
+            return fs.readFileSync(file, "utf-8");
+        } catch (e) {
+            this.logger.debug("Failed to read pyproject.toml", e);
+            return undefined;
+        }
+    }
+}