@oneuptime/common 10.8.1 → 11.0.0

package/build/dist/Server/Services/ProfileAggregationService.js

@@ -8,7 +8,9 @@ var __metadata = (this && this.__metadata) || function (k, v) {
     if (typeof Reflect === "object" && typeof Reflect.metadata === "function") return Reflect.metadata(k, v);
 };
 import { SQL } from "../Utils/AnalyticsDatabase/Statement";
+import { getQuerySettings } from "../Utils/AnalyticsDatabase/QuerySettingsHelper";
 import ProfileSampleDatabaseService from "./ProfileSampleService";
+import ProfileDatabaseService from "./ProfileService";
 import TableColumnType from "../../Types/AnalyticsDatabase/TableColumnType";
 import Includes from "../../Types/BaseDatabase/Includes";
 import AnalyticsTableName from "../../Types/AnalyticsDatabase/AnalyticsTableName";
@@ -20,17 +22,21 @@ export class ProfileAggregationService {
      *
      * Each sample has a `stacktrace` array where each element follows the
      * format "functionName@fileName:lineNumber".  The array is ordered
-     * bottom-up (index 0 = root, last index = leaf).
+     * LEAF-FIRST (index 0 = leaf, last index = root) — the pprof
+     * Sample.location_id and OTLP Stack.location_indices conventions both
+     * store the innermost frame first, and ingestion preserves wire order.
      *
-     * We aggregate samples that share common stack prefixes into a tree of
+     * Samples are pre-aggregated in ClickHouse by stack identity, then
+     * stacks sharing common prefixes are merged into a tree of
      * ProfileFlamegraphNode objects.
      */
     static async getFlamegraph(request) {
-        const statement = ProfileAggregationService.buildFlamegraphQuery(request);
+        const statement = ProfileAggregationService.buildGroupedStackQuery(request);
         const dbResult = await ProfileSampleDatabaseService.executeQuery(statement);
         const response = await dbResult.json();
         const rows = response.data || [];
-        // Build the tree from samples
+        const truncated = rows.length >= ProfileAggregationService.MAX_STACK_FETCH;
+        // Build the tree from the pre-aggregated stacks
         const root = {
             functionName: "(root)",
             fileName: "",
@@ -40,25 +46,37 @@ export class ProfileAggregationService {
             children: [],
             frameType: "",
         };
+        /*
+         * Per-node child index so frame lookup is O(1) instead of a linear
+         * children.find() — wide fan-out (thousands of siblings under root)
+         * made the scan quadratic. Kept outside the nodes so the serialized
+         * tree shape is unchanged.
+         */
+        const childIndex = new WeakMap();
         for (const row of rows) {
             const stacktrace = row["stacktrace"] || [];
             const frameTypes = row["frameTypes"] || [];
-            const value = Number(row["value"] || 0);
+            const value = Number(row["totalValue"] || 0);
             if (stacktrace.length === 0) {
                 continue;
             }
-            // Walk down the tree, creating nodes as needed
+            /*
+             * Walk down the tree, creating nodes as needed. Stored stacks are
+             * leaf-first, so iterate from the END of the array (root) toward
+             * index 0 (leaf) to build the tree root-to-leaf.
+             */
             let currentNode = root;
             currentNode.totalValue += value;
-            for (let i = 0; i < stacktrace.length; i++) {
+            for (let i = stacktrace.length - 1; i >= 0; i--) {
                 const frame = ProfileAggregationService.parseFrame(stacktrace[i]);
                 const frameType = frameTypes[i] || "";
-                // Find or create child
-                let childNode = currentNode.children.find((child) => {
-                    return (child.functionName === frame.functionName &&
-                        child.fileName === frame.fileName &&
-                        child.lineNumber === frame.lineNumber);
-                });
+                const frameKey = `${frame.functionName}@${frame.fileName}:${frame.lineNumber}`;
+                let index = childIndex.get(currentNode);
+                if (!index) {
+                    index = new Map();
+                    childIndex.set(currentNode, index);
+                }
+                let childNode = index.get(frameKey);
                 if (!childNode) {
                     childNode = {
                         functionName: frame.functionName,
@@ -70,33 +88,60 @@ export class ProfileAggregationService {
                         frameType: frameType,
                     };
                     currentNode.children.push(childNode);
+                    index.set(frameKey, childNode);
                 }
                 childNode.totalValue += value;
-                // If this is the leaf frame, add to selfValue
-                if (i === stacktrace.length - 1) {
+                // Index 0 is the leaf frame in leaf-first storage — it gets selfValue
+                if (i === 0) {
                     childNode.selfValue += value;
                 }
                 currentNode = childNode;
             }
         }
-        return root;
+        return { flamegraph: root, truncated };
     }
     /**
      * Return the top functions aggregated across samples, sorted by the
-     * requested metric (selfValue, totalValue, or sampleCount).
+     * requested metric (selfValue, totalValue, or sampleCount), along with
+     * the pre-limit window total so callers can show "% of total" figures.
      */
     static async getFunctionList(request) {
-        var _a;
-        const statement = ProfileAggregationService.buildFunctionListQuery(request);
-        const dbResult = await ProfileSampleDatabaseService.executeQuery(statement);
+        var _a, _b;
+        const filters = Object.assign(Object.assign(Object.assign(Object.assign(Object.assign(Object.assign({ projectId: request.projectId }, (request.startTime !== undefined && { startTime: request.startTime })), (request.endTime !== undefined && { endTime: request.endTime })), (request.profileId !== undefined && { profileId: request.profileId })), (request.serviceIds !== undefined && {
+            serviceIds: request.serviceIds,
+        })), (request.profileType !== undefined && {
+            profileType: request.profileType,
+        })), (request.profileTypes !== undefined && {
+            profileTypes: request.profileTypes,
+        }));
+        const statement = ProfileAggregationService.buildGroupedStackQuery(filters);
+        const windowTotalStatement = ProfileAggregationService.buildWindowTotalQuery(filters);
+        /*
+         * The window total must be computed over ALL matching rows (pre-limit),
+         * so it cannot be derived from the capped grouped result. It is a
+         * single cheap aggregate, so run both queries concurrently.
+         */
+        const [dbResult, windowTotalDbResult] = await Promise.all([
+            ProfileSampleDatabaseService.executeQuery(statement),
+            ProfileSampleDatabaseService.executeQuery(windowTotalStatement),
+        ]);
         const response = await dbResult.json();
+        const windowTotalResponse = await windowTotalDbResult.json();
         const rows = response.data || [];
-        // Aggregate per-function stats in-memory from the raw samples
+        const truncated = rows.length >= ProfileAggregationService.MAX_STACK_FETCH;
+        const windowTotalRows = windowTotalResponse.data || [];
+        const windowTotal = Number(((_a = windowTotalRows[0]) === null || _a === void 0 ? void 0 : _a["windowTotal"]) || 0);
+        // Aggregate per-function stats from the pre-aggregated stacks
         const functionMap = new Map();
         for (const row of rows) {
             const stacktrace = row["stacktrace"] || [];
             const frameTypes = row["frameTypes"] || [];
-            const value = Number(row["value"] || 0);
+            const value = Number(row["totalValue"] || 0);
+            /*
+             * Each grouped row stands in for `sampleCount` raw samples sharing
+             * one stack, so per-function sample counts add the group size, not 1.
+             */
+            const groupSampleCount = Number(row["sampleCount"] || 0);
             if (stacktrace.length === 0) {
                 continue;
             }
@@ -105,7 +150,8 @@ export class ProfileAggregationService {
                 const frame = ProfileAggregationService.parseFrame(stacktrace[i]);
                 const frameType = frameTypes[i] || "";
                 const key = `${frame.functionName}@${frame.fileName}:${frame.lineNumber}`;
-                const isLeaf = i === stacktrace.length - 1;
+                // Stacks are stored leaf-first, so index 0 is the leaf frame.
+                const isLeaf = i === 0;
                 let entry = functionMap.get(key);
                 if (!entry) {
                     entry = {
@@ -118,10 +164,10 @@ export class ProfileAggregationService {
                     };
                     functionMap.set(key, entry);
                 }
-                // totalValue: count the value once per unique function per sample
+                // totalValue: count the value once per unique function per stack
                 if (!seenInThisSample.has(key)) {
                     entry.totalValue += value;
-                    entry.sampleCount += 1;
+                    entry.sampleCount += groupSampleCount;
                     seenInThisSample.add(key);
                 }
                 // selfValue: only the leaf frame
@@ -142,29 +188,250 @@ export class ProfileAggregationService {
             }
             return b.selfValue - a.selfValue;
         });
-        const limit = (_a = request.limit) !== null && _a !== void 0 ? _a : ProfileAggregationService.DEFAULT_FUNCTION_LIST_LIMIT;
-        return items.slice(0, limit);
+        const limit = (_b = request.limit) !== null && _b !== void 0 ? _b : ProfileAggregationService.DEFAULT_FUNCTION_LIST_LIMIT;
+        return {
+            functions: items.slice(0, limit),
+            windowTotal,
+            truncated,
+        };
     }
     /**
-     * Build a diff flamegraph comparing two time ranges.
-     * Returns a tree where each node has baseline/comparison values and deltas.
+     * "Sandwich" aggregation for one function: every matching stack is
+     * split at the focused function, the frames toward the root become the
+     * callers tree and the frames toward the leaf become the callees tree.
+     * Both trees are rooted at the focused function itself so the UI can
+     * render them stacked around a single center row.
+     *
+     * Frames match on functionName + fileName only — line numbers shift on
+     * every deploy (and with inlining), so including them would split one
+     * logical function into many.
      */
-    static async getDiffFlamegraph(request) {
-        const baselineTree = await ProfileAggregationService.getFlamegraph(Object.assign(Object.assign(Object.assign({ projectId: request.projectId, startTime: request.baselineStartTime, endTime: request.baselineEndTime }, (request.serviceIds !== undefined && {
-            serviceIds: request.serviceIds,
-        })), (request.profileType !== undefined && {
-            profileType: request.profileType,
-        })), (request.profileTypes !== undefined && {
-            profileTypes: request.profileTypes,
-        })));
-        const comparisonTree = await ProfileAggregationService.getFlamegraph(Object.assign(Object.assign(Object.assign({ projectId: request.projectId, startTime: request.comparisonStartTime, endTime: request.comparisonEndTime }, (request.serviceIds !== undefined && {
+    static async getFunctionFocus(request) {
+        var _a;
+        const filters = Object.assign(Object.assign(Object.assign(Object.assign(Object.assign(Object.assign({ projectId: request.projectId }, (request.profileId !== undefined && { profileId: request.profileId })), (request.startTime !== undefined && { startTime: request.startTime })), (request.endTime !== undefined && { endTime: request.endTime })), (request.serviceIds !== undefined && {
             serviceIds: request.serviceIds,
         })), (request.profileType !== undefined && {
             profileType: request.profileType,
         })), (request.profileTypes !== undefined && {
             profileTypes: request.profileTypes,
-        })));
-        return ProfileAggregationService.mergeDiffTrees(baselineTree, comparisonTree);
+        }));
+        const statement = ProfileAggregationService.buildFunctionStackQuery(filters, request.functionName, request.fileName);
+        const windowTotalStatement = ProfileAggregationService.buildWindowTotalQuery(filters);
+        /*
+         * windowTotal deliberately ignores the function prefilter — it is the
+         * denominator for "% of window" figures, so it must cover every row
+         * matching the non-function filters. Both reads are independent, so
+         * run them concurrently.
+         */
+        const [dbResult, windowTotalDbResult] = await Promise.all([
+            ProfileSampleDatabaseService.executeQuery(statement),
+            ProfileSampleDatabaseService.executeQuery(windowTotalStatement),
+        ]);
+        const response = await dbResult.json();
+        const windowTotalResponse = await windowTotalDbResult.json();
+        const rows = response.data || [];
+        const truncated = rows.length >= ProfileAggregationService.MAX_STACK_FETCH;
+        const windowTotalRows = windowTotalResponse.data || [];
+        const windowTotal = Number(((_a = windowTotalRows[0]) === null || _a === void 0 ? void 0 : _a["windowTotal"]) || 0);
+        /*
+         * Both trees are rooted at the focused function. lineNumber is 0
+         * because the view aggregates across line numbers by design.
+         */
+        const callers = {
+            functionName: request.functionName,
+            fileName: request.fileName,
+            lineNumber: 0,
+            selfValue: 0,
+            totalValue: 0,
+            children: [],
+            frameType: "",
+        };
+        const callees = {
+            functionName: request.functionName,
+            fileName: request.fileName,
+            lineNumber: 0,
+            selfValue: 0,
+            totalValue: 0,
+            children: [],
+            frameType: "",
+        };
+        /*
+         * Per-node child index for O(1) frame lookup, mirroring getFlamegraph.
+         * One map serves both trees since it is keyed by parent node identity.
+         */
+        const childIndex = new WeakMap();
+        let totalValue = 0;
+        let selfValue = 0;
+        let sampleCount = 0;
+        for (const row of rows) {
+            const stacktrace = row["stacktrace"] || [];
+            const frameTypes = row["frameTypes"] || [];
+            const value = Number(row["totalValue"] || 0);
+            const groupSampleCount = Number(row["sampleCount"] || 0);
+            if (stacktrace.length === 0) {
+                continue;
+            }
+            /*
+             * Stacks are leaf-first, so scanning from the END finds the
+             * occurrence CLOSEST TO THE ROOT first. For recursive functions
+             * this attributes the whole recursive subtree (including nested
+             * occurrences) to one split point, so each stack's value is
+             * counted exactly once.
+             */
+            let occurrenceIndex = -1;
+            for (let i = stacktrace.length - 1; i >= 0; i--) {
+                const frame = ProfileAggregationService.parseFrame(stacktrace[i]);
+                if (frame.functionName === request.functionName &&
+                    frame.fileName === request.fileName) {
+                    occurrenceIndex = i;
+                    break;
+                }
+            }
+            /*
+             * The ClickHouse prefilter is a prefix match and can overmatch
+             * when a fileName containing colons shares a prefix with another
+             * (parseFrame splits on the LAST colon) — drop those rows here.
+             */
+            if (occurrenceIndex === -1) {
+                continue;
+            }
+            totalValue += value;
+            sampleCount += groupSampleCount;
+            /*
+             * Self time counts whenever the function is the EXECUTING (leaf)
+             * frame — checked against index 0 directly rather than the split
+             * point, because a recursive stack splits at the root-most
+             * occurrence while the leaf may still be the same function
+             * (standard sandwich-view semantics: self = "was on CPU").
+             */
+            const leafFrame = ProfileAggregationService.parseFrame(stacktrace[0]);
+            if (leafFrame.functionName === request.functionName &&
+                leafFrame.fileName === request.fileName) {
+                selfValue += value;
+            }
+            const occurrenceFrameType = frameTypes[occurrenceIndex] || "";
+            if (!callers.frameType) {
+                callers.frameType = occurrenceFrameType;
+            }
+            if (!callees.frameType) {
+                callees.frameType = occurrenceFrameType;
+            }
+            callers.totalValue += value;
+            callees.totalValue += value;
+            /*
+             * Caller chain: frames ABOVE the occurrence (toward the root),
+             * direct caller first so depth in the tree equals caller distance.
+             */
+            const callerIndices = [];
+            for (let i = occurrenceIndex + 1; i < stacktrace.length; i++) {
+                callerIndices.push(i);
+            }
+            /*
+             * Callee chain: frames BELOW the occurrence (toward the leaf),
+             * direct callee first.
+             */
+            const calleeIndices = [];
+            for (let i = occurrenceIndex - 1; i >= 0; i--) {
+                calleeIndices.push(i);
+            }
+            ProfileAggregationService.appendFocusChain({
+                root: callers,
+                stacktrace,
+                frameTypes,
+                frameIndices: callerIndices,
+                value,
+                childIndex,
+            });
+            ProfileAggregationService.appendFocusChain({
+                root: callees,
+                stacktrace,
+                frameTypes,
+                frameIndices: calleeIndices,
+                value,
+                childIndex,
+            });
+        }
+        return {
+            functionName: request.functionName,
+            fileName: request.fileName,
+            totalValue,
+            selfValue,
+            sampleCount,
+            windowTotal,
+            callers,
+            callees,
+            truncated,
+        };
+    }
+    /**
+     * Walk one split half of a stack into a focus tree, merging frames on
+     * functionName + fileName (line numbers aggregated away — same
+     * deploy-stability rationale as the focused-function match itself).
+     * The terminal node of the chain absorbs selfValue: in the callers
+     * tree that is the stack root ("the chain starts here"), in the
+     * callees tree the leaf ("the time bottoms out here"). An empty chain
+     * terminates at the focused root itself.
+     */
+    static appendFocusChain(data) {
+        let currentNode = data.root;
+        for (const frameIndex of data.frameIndices) {
+            const frame = ProfileAggregationService.parseFrame(data.stacktrace[frameIndex]);
+            const frameType = data.frameTypes[frameIndex] || "";
+            const frameKey = `${frame.functionName}@${frame.fileName}`;
+            let index = data.childIndex.get(currentNode);
+            if (!index) {
+                index = new Map();
+                data.childIndex.set(currentNode, index);
+            }
+            let childNode = index.get(frameKey);
+            if (!childNode) {
+                childNode = {
+                    functionName: frame.functionName,
+                    fileName: frame.fileName,
+                    lineNumber: 0,
+                    selfValue: 0,
+                    totalValue: 0,
+                    children: [],
+                    frameType: frameType,
+                };
+                currentNode.children.push(childNode);
+                index.set(frameKey, childNode);
+            }
+            childNode.totalValue += data.value;
+            currentNode = childNode;
+        }
+        currentNode.selfValue += data.value;
+    }
+    /**
+     * Build a diff flamegraph comparing two time ranges.
+     * Returns a tree where each node has baseline/comparison values and
+     * deltas, plus a truncation flag covering both source windows.
+     */
+    static async getDiffFlamegraph(request) {
+        /*
+         * The two windows are independent reads, so fetch them concurrently —
+         * sequential awaits doubled the latency of every diff request.
+         */
+        const [baselineResult, comparisonResult] = await Promise.all([
+            ProfileAggregationService.getFlamegraph(Object.assign(Object.assign(Object.assign({ projectId: request.projectId, startTime: request.baselineStartTime, endTime: request.baselineEndTime }, (request.serviceIds !== undefined && {
+                serviceIds: request.serviceIds,
+            })), (request.profileType !== undefined && {
+                profileType: request.profileType,
+            })), (request.profileTypes !== undefined && {
+                profileTypes: request.profileTypes,
+            }))),
+            ProfileAggregationService.getFlamegraph(Object.assign(Object.assign(Object.assign({ projectId: request.projectId, startTime: request.comparisonStartTime, endTime: request.comparisonEndTime }, (request.serviceIds !== undefined && {
+                serviceIds: request.serviceIds,
+            })), (request.profileType !== undefined && {
+                profileType: request.profileType,
+            })), (request.profileTypes !== undefined && {
+                profileTypes: request.profileTypes,
+            }))),
+        ]);
+        return {
+            diffFlamegraph: ProfileAggregationService.mergeDiffTrees(baselineResult.flamegraph, comparisonResult.flamegraph),
+            truncated: baselineResult.truncated || comparisonResult.truncated,
+        };
     }
     static mergeDiffTrees(baseline, comparison) {
         const baselineValue = (baseline === null || baseline === void 0 ? void 0 : baseline.totalValue) || 0;
@@ -221,134 +488,239 @@ export class ProfileAggregationService {
         return node;
     }
     /**
-     * Aggregate sample / profile counts per serviceId for a time window.
+     * Aggregate sample / profile counts per primaryEntityId for a time window.
      * Drives the "loudest services first" sort on the Profiles dashboard
      * so a developer opening the page lands on the workloads that are
      * actually doing work rather than scrolling past kernel-thread noise.
+     *
+     * Reads the small Profile table (one row per ingested profile, with a
+     * denormalized sampleCount) instead of scanning every row of the huge
+     * ProfileSample table for what is just a per-service ranking.
      */
     static async getServiceActivity(request) {
         const statement = ProfileAggregationService.buildServiceActivityQuery(request);
-        const dbResult = await ProfileSampleDatabaseService.executeQuery(statement);
+        const dbResult = await ProfileDatabaseService.executeQuery(statement);
         const response = await dbResult.json();
         const rows = response.data || [];
         const out = [];
         for (const row of rows) {
-            const serviceId = String(row["serviceId"] || "");
-            if (!serviceId) {
+            const primaryEntityId = String(row["primaryEntityId"] || "");
+            if (!primaryEntityId) {
                 continue;
             }
             out.push({
-                serviceId,
-                sampleCount: Number(row["sampleCount"] || 0),
+                primaryEntityId,
+                sampleCount: Number(row["totalSampleCount"] || 0),
                 profileCount: Number(row["profileCount"] || 0),
-                totalValue: Number(row["totalValue"] || 0),
+                /*
+                 * Profile rows do not carry a summed sample value, and no client
+                 * reads this field — it is kept at 0 purely so the response shape
+                 * stays stable for existing consumers.
+                 */
+                totalValue: 0,
             });
         }
         return out;
     }
+    /**
+     * Group profiling volume by a dimension ("service" = primaryEntityId,
+     * anything else = a Profile attribute key) over a time window.
+     *
+     * Reads the small Profile table (one row per ingested profile with a
+     * denormalized sampleCount) — same reasoning as getServiceActivity: a
+     * share-of-volume ranking never needs the huge ProfileSample table.
+     *
+     * totalSampleCount sums over ALL fetched groups before the limit is
+     * applied, so item shares stay percentages of the real total rather
+     * than of the visible top-N.
+     */
+    static async getBreakdown(request) {
+        var _a;
+        const statement = ProfileAggregationService.buildBreakdownQuery(request);
+        const dbResult = await ProfileDatabaseService.executeQuery(statement);
+        const response = await dbResult.json();
+        const rows = response.data || [];
+        // Rows arrive sorted by summed sampleCount descending (query ORDER BY).
+        const allItems = [];
+        let totalSampleCount = 0;
+        for (const row of rows) {
+            const value = String(row["breakdownValue"] || "");
+            if (!value) {
+                continue;
+            }
+            const groupSampleCount = Number(row["totalSampleCount"] || 0);
+            totalSampleCount += groupSampleCount;
+            allItems.push({
+                value,
+                sampleCount: groupSampleCount,
+                profileCount: Number(row["profileCount"] || 0),
+                share: 0,
+            });
+        }
+        const limit = (_a = request.limit) !== null && _a !== void 0 ? _a : ProfileAggregationService.DEFAULT_BREAKDOWN_LIMIT;
+        const items = allItems.slice(0, limit);
+        for (const item of items) {
+            item.share =
+                totalSampleCount > 0 ? (item.sampleCount / totalSampleCount) * 100 : 0;
+        }
+        return { items, totalSampleCount };
+    }
     // --- Query builders ---
-    static buildFlamegraphQuery(request) {
+    /**
+     * Fetch samples pre-aggregated by stack identity. Collapsing identical
+     * (stacktrace, frameTypes) pairs in ClickHouse shrinks tens of thousands
+     * of raw sample rows to a few hundred unique stacks before they cross
+     * the wire, and the ORDER BY makes the LIMIT drop the smallest stacks
+     * first instead of an arbitrary subset.
+     *
+     * `value` is stored as Int128 in ClickHouse, so the sum is wrapped in
+     * toFloat64 to keep the JSON output numeric (toFloat64OrZero would fail
+     * with "Illegal type Int128").
+     */
+    static buildGroupedStackQuery(request) {
         const statement = SQL `
       SELECT
         stacktrace,
         frameTypes,
-        value
+        toFloat64(sum(value)) AS totalValue,
+        count() AS sampleCount
       FROM ${ProfileAggregationService.TABLE_NAME}
       WHERE projectId = ${{
             type: TableColumnType.ObjectID,
             value: request.projectId,
         }}
     `;
-        if (request.profileId) {
-            statement.append(SQL ` AND profileId = ${{
-                type: TableColumnType.Text,
-                value: request.profileId,
-            }}`);
-        }
-        if (request.startTime) {
-            statement.append(SQL ` AND time >= ${{
-                type: TableColumnType.Date,
-                value: request.startTime,
-            }}`);
-        }
-        if (request.endTime) {
-            statement.append(SQL ` AND time <= ${{
-                type: TableColumnType.Date,
-                value: request.endTime,
-            }}`);
-        }
-        ProfileAggregationService.appendCommonFilters(statement, request);
-        statement.append(SQL ` LIMIT ${{
+        ProfileAggregationService.appendSampleFilters(statement, request);
+        statement.append(SQL ` GROUP BY stacktrace, frameTypes
+           ORDER BY totalValue DESC
+           LIMIT ${{
             type: TableColumnType.Number,
-            value: ProfileAggregationService.MAX_SAMPLE_FETCH,
+            value: ProfileAggregationService.MAX_STACK_FETCH,
         }}`);
         /*
-         * Cap runtime below the client's 58s request_timeout; a flamegraph
-         * over a busy project can pull MAX_SAMPLE_FETCH raw samples. 'break'
-         * yields a partial flamegraph rather than holding a pool connection.
+         * Cap runtime below the client's 58s request_timeout; 'break' yields
+         * a partial (top stacks) result rather than holding a pool connection.
          */
-        statement.append(" SETTINGS max_execution_time = 45, timeout_overflow_mode = 'break'");
+        statement.append(getQuerySettings({
+            maxExecutionTimeInSeconds: 45,
+            timeoutOverflowMode: "break",
+        }));
         return statement;
     }
-    static buildFunctionListQuery(request) {
+    /**
+     * Grouped-stack query restricted to stacks containing one function, so
+     * a sandwich aggregation never pulls the full window's stacks over the
+     * wire just to discard most of them.
+     *
+     * Stored frame strings are "functionName" + ("@fileName" when fileName
+     * is non-empty) + (":lineNumber" when lineNumber > 0) — see
+     * OtelProfilesIngestService.resolveStackFrames. Line numbers shift
+     * across deploys, so the predicate matches the functionName+fileName
+     * identity only:
+     *   - non-empty fileName: exact "fn@file" (no line recorded) OR prefix
+     *     "fn@file:" (line recorded);
+     *   - empty fileName: the exact bare frame (folded uploads and
+     *     address-only frames store just the name — parseFrame treats any
+     *     '@'-less frame as all functionName, so equality is the full
+     *     match).
+     * The prefix form can overmatch when one colon-bearing fileName is a
+     * prefix of another (parseFrame splits on the LAST colon), so callers
+     * must re-check each frame in memory via parseFrame.
+     */
+    static buildFunctionStackQuery(request, functionName, fileName) {
         const statement = SQL `
       SELECT
         stacktrace,
         frameTypes,
-        value
+        toFloat64(sum(value)) AS totalValue,
+        count() AS sampleCount
       FROM ${ProfileAggregationService.TABLE_NAME}
       WHERE projectId = ${{
             type: TableColumnType.ObjectID,
             value: request.projectId,
         }}
-        AND time >= ${{
-            type: TableColumnType.Date,
-            value: request.startTime,
-        }}
-        AND time <= ${{
-            type: TableColumnType.Date,
-            value: request.endTime,
-        }}
     `;
-        ProfileAggregationService.appendCommonFilters(statement, request);
-        statement.append(SQL ` LIMIT ${{
+        ProfileAggregationService.appendSampleFilters(statement, request);
+        if (fileName) {
+            statement.append(SQL ` AND arrayExists(x -> (x = ${{
+                type: TableColumnType.Text,
+                value: `${functionName}@${fileName}`,
+            }} OR startsWith(x, ${{
+                type: TableColumnType.Text,
+                value: `${functionName}@${fileName}:`,
+            }})), stacktrace)`);
+        }
+        else {
+            statement.append(SQL ` AND has(stacktrace, ${{
+                type: TableColumnType.Text,
+                value: functionName,
+            }})`);
+        }
+        statement.append(SQL ` GROUP BY stacktrace, frameTypes
+           ORDER BY totalValue DESC
+           LIMIT ${{
             type: TableColumnType.Number,
-            value: ProfileAggregationService.MAX_SAMPLE_FETCH,
+            value: ProfileAggregationService.MAX_STACK_FETCH,
         }}`);
         /*
          * Cap runtime below the client's 58s request_timeout; 'break' yields
-         * a partial function list rather than holding a pool connection.
+         * a partial (top stacks) result rather than holding a pool connection.
          */
-        statement.append(" SETTINGS max_execution_time = 45, timeout_overflow_mode = 'break'");
+        statement.append(getQuerySettings({
+            maxExecutionTimeInSeconds: 45,
+            timeoutOverflowMode: "break",
+        }));
+        return statement;
+    }
+    /**
+     * Total of `value` across every sample row matching the filters — no
+     * GROUP BY and no LIMIT, so it stays correct even when the grouped
+     * stack query truncates.
+     */
+    static buildWindowTotalQuery(request) {
+        const statement = SQL `
+      SELECT
+        toFloat64(sum(value)) AS windowTotal
+      FROM ${ProfileAggregationService.TABLE_NAME}
+      WHERE projectId = ${{
+            type: TableColumnType.ObjectID,
+            value: request.projectId,
+        }}
+    `;
+        ProfileAggregationService.appendSampleFilters(statement, request);
+        statement.append(getQuerySettings({
+            maxExecutionTimeInSeconds: 45,
+            timeoutOverflowMode: "break",
+        }));
         return statement;
     }
     static buildServiceActivityQuery(request) {
         /*
-         * sum(value) directly — value is stored as Int128 in ClickHouse, so
-         * toFloat64OrZero would fail ("Illegal type Int128"); a plain sum is
-         * the right idiom here, and we coerce to Number when serialising in
-         * getServiceActivity.
+         * One Profile row per ingested profile with a denormalized sampleCount,
+         * so the per-service ranking never has to touch the (orders of
+         * magnitude larger) ProfileSample table. uniqExact tolerates duplicate
+         * rows for the same profileId from re-ingestion.
          */
         const statement = SQL `
       SELECT
-        toString(serviceId) AS serviceId,
-        count() AS sampleCount,
-        uniqExact(profileId) AS profileCount,
-        toFloat64(sum(value)) AS totalValue
-      FROM ${ProfileAggregationService.TABLE_NAME}
+        toString(primaryEntityId) AS primaryEntityId,
+        toFloat64(sum(sampleCount)) AS totalSampleCount,
+        uniqExact(profileId) AS profileCount
+      FROM ${ProfileAggregationService.PROFILE_TABLE_NAME}
       WHERE projectId = ${{
             type: TableColumnType.ObjectID,
             value: request.projectId,
         }}
-        AND time >= ${{
+        AND startTime >= ${{
             type: TableColumnType.Date,
             value: request.startTime,
         }}
-        AND time <= ${{
+        AND startTime <= ${{
             type: TableColumnType.Date,
             value: request.endTime,
         }}
     `;
+        statement.append(" AND retentionDate >= now()");
         /*
          * profileTypes (array) wins over profileType (single) so the UI
          * can OR together every raw type string in a category.
@@ -365,19 +737,127 @@ export class ProfileAggregationService {
                 value: request.profileType,
             }}`);
         }
-        statement.append(SQL ` GROUP BY serviceId
-           ORDER BY sampleCount DESC
+        statement.append(SQL ` GROUP BY primaryEntityId
+           ORDER BY totalSampleCount DESC
            LIMIT 10000`);
         /*
          * Cap runtime below the client's 58s request_timeout; 'break' yields
          * partial service activity rather than holding a pool connection.
          */
-        statement.append(" SETTINGS max_execution_time = 45, timeout_overflow_mode = 'break'");
+        statement.append(getQuerySettings({
+            maxExecutionTimeInSeconds: 45,
+            timeoutOverflowMode: "break",
+        }));
         return statement;
     }
+    /**
+     * Per-group volume aggregate over the Profile table for getBreakdown.
+     *
+     * The "service" dimension groups by primaryEntityId (mirroring
+     * buildServiceActivityQuery); any other dimension is a Profile
+     * attribute key, read via the parameter-bound Map subscript
+     * `attributes[key]` — the same O(1) fast path the analytics query
+     * generator uses for map equality. The subscript yields '' for missing
+     * keys, so the != '' guard drops profiles without the attribute.
+     */
+    static buildBreakdownQuery(request) {
+        let statement;
+        if (request.breakdownBy === "service") {
+            statement = SQL `
+        SELECT
+          toString(primaryEntityId) AS breakdownValue,
+          toFloat64(sum(sampleCount)) AS totalSampleCount,
+          uniqExact(profileId) AS profileCount
+        FROM ${ProfileAggregationService.PROFILE_TABLE_NAME}
+        WHERE projectId = ${{
+                type: TableColumnType.ObjectID,
+                value: request.projectId,
+            }}
+      `;
+        }
+        else {
+            statement = SQL `
+        SELECT
+          attributes[${{
+                type: TableColumnType.Text,
+                value: request.breakdownBy,
+            }}] AS breakdownValue,
+          toFloat64(sum(sampleCount)) AS totalSampleCount,
+          uniqExact(profileId) AS profileCount
+        FROM ${ProfileAggregationService.PROFILE_TABLE_NAME}
+        WHERE projectId = ${{
+                type: TableColumnType.ObjectID,
+                value: request.projectId,
+            }}
+          AND attributes[${{
+                type: TableColumnType.Text,
+                value: request.breakdownBy,
+            }}] != ''
+      `;
+        }
+        statement.append(SQL ` AND startTime >= ${{
+            type: TableColumnType.Date,
+            value: request.startTime,
+        }} AND startTime <= ${{
+            type: TableColumnType.Date,
+            value: request.endTime,
+        }}`);
+        /*
+         * Read-side retention filter: rows past their per-service retention
+         * stay in their part until the whole part drops (ttl_only_drop_parts).
+         */
+        statement.append(" AND retentionDate >= now()");
+        ProfileAggregationService.appendCommonFilters(statement, request);
+        statement.append(SQL ` GROUP BY breakdownValue
+           ORDER BY totalSampleCount DESC
+           LIMIT ${{
+            type: TableColumnType.Number,
+            value: ProfileAggregationService.MAX_BREAKDOWN_GROUP_FETCH,
+        }}`);
+        /*
+         * Cap runtime below the client's 58s request_timeout; 'break' yields
+         * a partial (top groups) result rather than holding a pool connection.
+         */
+        statement.append(getQuerySettings({
+            maxExecutionTimeInSeconds: 45,
+            timeoutOverflowMode: "break",
+        }));
+        return statement;
+    }
+    /**
+     * Shared WHERE-clause tail for ProfileSample reads: optional profileId /
+     * time-window filters, the read-side retention filter, and the common
+     * service / profile-type filters.
+     */
+    static appendSampleFilters(statement, request) {
+        if (request.profileId) {
+            statement.append(SQL ` AND profileId = ${{
+                type: TableColumnType.Text,
+                value: request.profileId,
+            }}`);
+        }
+        if (request.startTime) {
+            statement.append(SQL ` AND time >= ${{
+                type: TableColumnType.Date,
+                value: request.startTime,
+            }}`);
+        }
+        if (request.endTime) {
+            statement.append(SQL ` AND time <= ${{
+                type: TableColumnType.Date,
+                value: request.endTime,
+            }}`);
+        }
+        /*
+         * Read-side retention filter: rows past their per-service retention
+         * stay in their part until the whole part drops (ttl_only_drop_parts).
+         */
+        statement.append(" AND retentionDate >= now()");
+        ProfileAggregationService.appendCommonFilters(statement, request);
+    }
     static appendCommonFilters(statement, request) {
         if (request.serviceIds && request.serviceIds.length > 0) {
-            statement.append(SQL ` AND serviceId IN (${{
+            statement.append(SQL ` AND primaryEntityId IN (${{
                 type: TableColumnType.ObjectID,
                 value: new Includes(request.serviceIds.map((id) => {
                     return id.toString();
@@ -436,8 +916,24 @@ export class ProfileAggregationService {
     }
 }
 ProfileAggregationService.TABLE_NAME = AnalyticsTableName.ProfileSample;
+ProfileAggregationService.PROFILE_TABLE_NAME = AnalyticsTableName.Profile;
 ProfileAggregationService.DEFAULT_FUNCTION_LIST_LIMIT = 50;
-ProfileAggregationService.MAX_SAMPLE_FETCH = 50000;
+ProfileAggregationService.DEFAULT_BREAKDOWN_LIMIT = 10;
+/**
+ * Cap on unique groups fetched by a breakdown query. Groups are ordered
+ * by summed sampleCount, so hitting the cap drops the quietest groups
+ * first — totals over the fetched groups stay representative even for
+ * pathological high-cardinality attribute keys.
+ */
+ProfileAggregationService.MAX_BREAKDOWN_GROUP_FETCH = 10000;
+/**
+ * Cap on unique (stacktrace, frameTypes) groups fetched per query.
+ * Sample reads GROUP BY stack identity and ORDER BY summed value, so
+ * hitting this cap drops the *smallest* stacks first — truncation is
+ * deterministic and barely visible in a flamegraph, unlike a LIMIT over
+ * raw unordered sample rows.
+ */
+ProfileAggregationService.MAX_STACK_FETCH = 10000;
 __decorate([
     CaptureSpan(),
     __metadata("design:type", Function),
@@ -450,6 +946,12 @@ __decorate([
     __metadata("design:paramtypes", [Object]),
     __metadata("design:returntype", Promise)
 ], ProfileAggregationService, "getFunctionList", null);
+__decorate([
+    CaptureSpan(),
+    __metadata("design:type", Function),
+    __metadata("design:paramtypes", [Object]),
+    __metadata("design:returntype", Promise)
+], ProfileAggregationService, "getFunctionFocus", null);
 __decorate([
     CaptureSpan(),
     __metadata("design:type", Function),
@@ -462,5 +964,11 @@ __decorate([
     __metadata("design:paramtypes", [Object]),
     __metadata("design:returntype", Promise)
 ], ProfileAggregationService, "getServiceActivity", null);
+__decorate([
+    CaptureSpan(),
+    __metadata("design:type", Function),
+    __metadata("design:paramtypes", [Object]),
+    __metadata("design:returntype", Promise)
+], ProfileAggregationService, "getBreakdown", null);
 export default ProfileAggregationService;
 //# sourceMappingURL=ProfileAggregationService.js.map