fix(webapp): keep the rest of a batch when one run or span has un-ingestable JSON

When a run's output/payload or a trace span's attributes contained JSON
ClickHouse could not ingest (e.g. an object nested past its depth limit),
the analytics insert aborted the whole batch. Because runs and events are
written in per-batch inserts, one bad row made every other run or span in
that batch go missing from the runs list, traces, and logs, even though the
Postgres rows were fine.

Both ClickHouse write paths now fall back to retrying the insert with
input_format_allow_errors so ClickHouse skips only the rows it can't parse
and commits the rest. The healthy insert path is unchanged; a successful
row isolation logs at info, only a failed isolation insert is an error.

Author: ericallam

.server-changes/fix-runs-list-batch-drop-on-unparseable-json.md

@@ -0,0 +1,6 @@
+---
+area: webapp
+type: fix
+---
+
+Fixed a rare case where a single run or span carrying data that could not be ingested would make other runs or trace events in the same batch go missing from the runs list, traces, and logs. Now only the affected item is dropped and everything else is stored normally.

apps/webapp/app/services/runsReplicationService.server.ts

@@ -1,6 +1,7 @@
 import type { ClickhouseFactory } from "~/services/clickhouse/clickhouseFactory.server";
 import {
   type ClickHouse,
+  type ClickHouseSettings,
   type PayloadInsertArray,
   type TaskRunInsertArray,
   composeTaskRunVersion,
@@ -175,14 +176,31 @@ export class RunsReplicationService {
   private _disableErrorFingerprinting: boolean;
 
   /**
-   * Counts batches that hit a ClickHouse `Cannot parse JSON object` failure
-   * that survived one sanitize-retry. These batches are dropped on the floor
-   * (returning success-ish to the caller so the retry layer doesn't spin on
-   * the same deterministic failure), and we track the drop count for
-   * observability. Counter only — does not gate behaviour.
+   * Counts batches dropped whole because even the row-isolation retry failed
+   * to insert anything (an unexpected terminal case). The common recovery now
+   * isolates the un-ingestable rows and lands the rest, so this should stay at
+   * zero in practice. Counter only — does not gate behaviour.
    */
   private _permanentlyDroppedBatches = 0;
 
+  /**
+   * Counts batches that took the row-isolation recovery path — a
+   * `Cannot parse JSON object` failure the sanitizer could not repair, where we
+   * re-inserted with `input_format_allow_errors_*` so ClickHouse skipped the
+   * un-ingestable rows and landed the rest. Reliable per-event signal that user
+   * data is hitting the ceiling.
+   */
+  private _rowIsolationRecoveries = 0;
+
+  /**
+   * Best-effort count of individual rows ClickHouse skipped across those
+   * row-isolation retries, derived from the insert summary's `written_rows`
+   * when the client surfaces it (it is not always populated under concurrent
+   * inserts, so treat this as a lower bound; `_rowIsolationRecoveries` is the
+   * dependable event count).
+   */
+  private _permanentlyDroppedRows = 0;
+
   // Metrics
   private _replicationLagHistogram: Histogram;
   private _batchesFlushedCounter: Counter;
@@ -416,6 +434,16 @@ export class RunsReplicationService {
     return this._permanentlyDroppedBatches;
   }
 
+  /** Exposed for tests and metrics — batches that took the row-isolation recovery path. */
+  get rowIsolationRecoveries() {
+    return this._rowIsolationRecoveries;
+  }
+
+  /** Exposed for tests and metrics — best-effort count of individual rows ClickHouse skipped during row-isolation retries. */
+  get permanentlyDroppedRows() {
+    return this._permanentlyDroppedRows;
+  }
+
   public async shutdown() {
     if (this._isShuttingDown) return;
 
@@ -870,15 +898,11 @@ export class RunsReplicationService {
           payloadError = plErr;
         }
 
-        // Only count rows that actually landed in ClickHouse. `kind: "dropped"`
-        // means the recovery wrapper bailed (sanitizer no-op or sanitize-retry
-        // still failed) — those rows never made it, so they must not show up
-        // as successful inserts in the per-batch counter.
-        if (!trErr && trOutcome?.kind !== "dropped") {
-          this._taskRunsInsertedCounter.add(group.taskRunInserts.length);
+        if (!trErr && trOutcome && trOutcome.kind !== "dropped") {
+          this._taskRunsInsertedCounter.add(landedRowCount(group.taskRunInserts.length, trOutcome));
         }
-        if (!plErr && plOutcome?.kind !== "dropped") {
-          this._payloadsInsertedCounter.add(group.payloadInserts.length);
+        if (!plErr && plOutcome && plOutcome.kind !== "dropped") {
+          this._payloadsInsertedCounter.add(landedRowCount(group.payloadInserts.length, plOutcome));
         }
       }
 
@@ -1034,10 +1058,14 @@ export class RunsReplicationService {
       return;
     }
     return await startSpan(this._tracer, "insertTaskRunsInserts", async (span) => {
-      const doInsert = async () => {
+      const doInsert = async (extraSettings?: ClickHouseSettings) => {
         const [insertError, insertResult] = await clickhouse.taskRuns.insertCompactArrays(
           taskRunInserts,
-          { params: { clickhouse_settings: this.#getClickhouseInsertSettings() } }
+          {
+            params: {
+              clickhouse_settings: { ...this.#getClickhouseInsertSettings(), ...extraSettings },
+            },
+          }
         );
         if (insertError) {
           this.logger.error("Error inserting task run inserts attempt", {
@@ -1068,10 +1096,14 @@ export class RunsReplicationService {
       return;
     }
     return await startSpan(this._tracer, "insertPayloadInserts", async (span) => {
-      const doInsert = async () => {
+      const doInsert = async (extraSettings?: ClickHouseSettings) => {
         const [insertError, insertResult] = await clickhouse.taskRuns.insertPayloadsCompactArrays(
           payloadInserts,
-          { params: { clickhouse_settings: this.#getClickhouseInsertSettings() } }
+          {
+            params: {
+              clickhouse_settings: { ...this.#getClickhouseInsertSettings(), ...extraSettings },
+            },
+          }
         );
         if (insertError) {
           this.logger.error("Error inserting payload inserts attempt", {
@@ -1094,102 +1126,132 @@ export class RunsReplicationService {
   }
 
   /**
-   * Wraps a ClickHouse insert with reactive UTF-16 sanitization for
-   * `Cannot parse JSON object` rejections. Mirrors the pattern from
-   * `ClickhouseEventRepository.#insertWithJsonParseRecovery` introduced
-   * in #3659 — same root cause (lone UTF-16 surrogates in user-provided
-   * JSON), same recovery shape:
+   * Wraps a ClickHouse insert with recovery for `Cannot parse JSON object`
+   * rejections so one un-ingestable row never drops its whole batch. Builds on
+   * the sanitize pattern from `ClickhouseEventRepository.#insertWithJsonParseRecovery`
+   * (#3659) and adds a final row-isolation fallback:
    *
-   *   1. Try the insert. Healthy batches pay zero scan cost.
-   *   2. On parse error, walk the whole batch via `sanitizeRows` and
-   *      replace any lone-surrogate string with `"[invalid-utf16]"`.
-   *   3. Retry once. If the sanitizer found nothing or the retry also
-   *      fails with the same error class, drop the batch loudly and
-   *      return — do NOT rethrow, otherwise the surrounding
-   *      `#insertWithRetry` layer would spin three more times on the
-   *      same deterministic failure.
-   *   4. Non-parse errors propagate unchanged so the existing
-   *      transient-retry path still handles them.
-   *
-   * The whole-batch scan (rather than slicing on the `at row N` hint) is
-   * deliberate: `at row N` semantics under `input_format_parallel_parsing`
-   * aren't stable enough to safely skip rows. The cost is bounded because
-   * `detectBadJsonStrings` exits in O(1) for clean strings.
+   *   1. Try the insert. Healthy batches pay zero recovery cost.
+   *   2. On a parse error, walk the batch via `sanitizeRows` and losslessly
+   *      repair what we can in place (lone UTF-16 surrogates → `"[invalid-utf16]"`,
+   *      out-of-range integers → their string form), then retry once. A clean
+   *      retry keeps every row.
+   *   3. If the sanitizer found nothing to fix, or the retry still hits the same
+   *      parse error (a structurally un-ingestable row, e.g. JSON nested past
+   *      ClickHouse's `input_format_json_max_depth`), retry once more with
+   *      `input_format_allow_errors_*` so ClickHouse skips only the un-parseable
+   *      rows and lands the rest. Verified against ClickHouse 26.2 (the cloud
+   *      version): the good rows land, the JSONCompactEachRowWithNames header is
+   *      unaffected, both sync and async.
+   *   4. Only if that row-isolation insert itself fails do we drop the batch,
+   *      and never rethrow a parse error — otherwise the surrounding
+   *      `#insertWithRetry` layer would spin on the same deterministic failure.
+   *   5. Non-parse errors propagate unchanged so the transient-retry path still
+   *      handles them.
    */
   async #insertWithJsonParseRecovery<T extends object>(
     rows: T[],
-    doInsert: () => Promise<unknown>,
+    doInsert: (extraSettings?: ClickHouseSettings) => Promise<unknown>,
     contextLabel: string,
     attempt: number
   ): Promise<
     | { kind: "inserted"; insertResult: unknown }
     | { kind: "sanitized"; insertResult: unknown }
+    | { kind: "isolated"; insertResult: unknown; rowsSkipped: number | null }
     | { kind: "dropped" }
   > {
     try {
       return { kind: "inserted", insertResult: await doInsert() };
     } catch (firstError) {
       if (!isClickHouseJsonParseError(firstError)) throw firstError;
 
-      const firstMessage =
-        typeof firstError === "object" && firstError !== null && "message" in firstError
-          ? String((firstError as { message?: unknown }).message ?? "")
-          : String(firstError);
-
+      const firstMessage = errorMessage(firstError);
       const rowHint = parseRowNumberFromError(firstMessage);
       const { rowsTouched, fieldsSanitized } = sanitizeRows(rows);
 
-      if (fieldsSanitized === 0) {
-        this._permanentlyDroppedBatches += 1;
-        this.logger.error(
-          "Dropped batch — ClickHouse JSON parse error but sanitizer found nothing to fix",
-          {
-            contextLabel,
-            attempt,
-            batchSize: rows.length,
-            clickhouseRowHint: rowHint,
-            permanentlyDroppedBatches: this._permanentlyDroppedBatches,
-            sampleRow: JSON.stringify(rows[0] ?? null).slice(0, 1024),
-            clickhouseError: firstMessage.split("\n")[0],
-          }
-        );
-        return { kind: "dropped" };
+      if (fieldsSanitized > 0) {
+        this.logger.warn("Sanitizing batch after ClickHouse JSON parse error", {
+          contextLabel,
+          attempt,
+          batchSize: rows.length,
+          clickhouseRowHint: rowHint,
+          rowsTouched,
+          fieldsSanitized,
+          clickhouseError: firstMessage.split("\n")[0],
+        });
+
+        try {
+          return { kind: "sanitized", insertResult: await doInsert() };
+        } catch (retryError) {
+          if (!isClickHouseJsonParseError(retryError)) throw retryError;
+        }
       }
 
-      this.logger.warn("Sanitizing batch after ClickHouse JSON parse error", {
-        contextLabel,
-        attempt,
-        batchSize: rows.length,
-        clickhouseRowHint: rowHint,
-        rowsTouched,
-        fieldsSanitized,
-        clickhouseError: firstMessage.split("\n")[0],
+      return await this.#insertIsolatingBadRows(rows, doInsert, contextLabel, attempt, firstMessage);
+    }
+  }
+
+  /**
+   * Last-resort recovery: re-run the insert with `input_format_allow_errors_*`
+   * so ClickHouse skips the rows it cannot parse and lands the rest, rather than
+   * aborting the whole INSERT. Forces a synchronous insert so `written_rows`
+   * reflects exactly what landed and we can count what was lost.
+   */
+  async #insertIsolatingBadRows<T extends object>(
+    rows: T[],
+    doInsert: (extraSettings?: ClickHouseSettings) => Promise<unknown>,
+    contextLabel: string,
+    attempt: number,
+    firstMessage: string
+  ): Promise<
+    { kind: "isolated"; insertResult: unknown; rowsSkipped: number | null } | { kind: "dropped" }
+  > {
+    try {
+      const insertResult = await doInsert({
+        input_format_allow_errors_num: String(rows.length),
+        input_format_allow_errors_ratio: 1,
+        async_insert: 0,
       });
 
-      try {
-        return { kind: "sanitized", insertResult: await doInsert() };
-      } catch (retryError) {
-        if (!isClickHouseJsonParseError(retryError)) throw retryError;
-
-        this._permanentlyDroppedBatches += 1;
-        const retryMessage =
-          typeof retryError === "object" && retryError !== null && "message" in retryError
-            ? String((retryError as { message?: unknown }).message ?? "")
-            : String(retryError);
-        this.logger.error(
-          "Dropped batch after sanitize-retry still hit ClickHouse JSON parse error",
-          {
-            contextLabel,
-            attempt,
-            batchSize: rows.length,
-            permanentlyDroppedBatches: this._permanentlyDroppedBatches,
-            sampleRow: JSON.stringify(rows[0] ?? null).slice(0, 1024),
-            firstError: firstMessage.split("\n")[0],
-            retryError: retryMessage.split("\n")[0],
-          }
-        );
-        return { kind: "dropped" };
+      this._rowIsolationRecoveries += 1;
+      const writtenRows = extractWrittenRows(insertResult);
+      const rowsSkipped = writtenRows === null ? null : Math.max(0, rows.length - writtenRows);
+      if (rowsSkipped !== null) {
+        this._permanentlyDroppedRows += rowsSkipped;
       }
+
+      this.logger.info(
+        "Isolated un-ingestable rows after ClickHouse JSON parse error — landed the rest of the batch",
+        {
+          contextLabel,
+          attempt,
+          batchSize: rows.length,
+          rowsSkipped,
+          rowIsolationRecoveries: this._rowIsolationRecoveries,
+          permanentlyDroppedRows: this._permanentlyDroppedRows,
+          sampleRow: JSON.stringify(rows[0] ?? null).slice(0, 1024),
+          clickhouseError: firstMessage.split("\n")[0],
+        }
+      );
+
+      return { kind: "isolated", insertResult, rowsSkipped };
+    } catch (isolationError) {
+      if (!isClickHouseJsonParseError(isolationError)) throw isolationError;
+
+      this._permanentlyDroppedBatches += 1;
+      this.logger.error(
+        "Dropped batch — row-isolation insert failed after ClickHouse JSON parse error",
+        {
+          contextLabel,
+          attempt,
+          batchSize: rows.length,
+          permanentlyDroppedBatches: this._permanentlyDroppedBatches,
+          sampleRow: JSON.stringify(rows[0] ?? null).slice(0, 1024),
+          firstError: firstMessage.split("\n")[0],
+          isolationError: errorMessage(isolationError).split("\n")[0],
+        }
+      );
+      return { kind: "dropped" };
     }
   }
 
@@ -1542,3 +1604,27 @@ function lsnToUInt64(lsn: string): bigint {
   const [seg, off] = lsn.split("/");
   return (BigInt("0x" + seg) << 32n) | BigInt("0x" + off);
 }
+
+function errorMessage(err: unknown): string {
+  return typeof err === "object" && err !== null && "message" in err
+    ? String((err as { message?: unknown }).message ?? "")
+    : String(err);
+}
+
+function landedRowCount(
+  groupSize: number,
+  outcome: { kind: string; rowsSkipped?: number | null }
+): number {
+  if (outcome.kind === "isolated" && typeof outcome.rowsSkipped === "number") {
+    return Math.max(0, groupSize - outcome.rowsSkipped);
+  }
+  return groupSize;
+}
+
+function extractWrittenRows(insertResult: unknown): number | null {
+  const written = (insertResult as { summary?: { written_rows?: string } } | null | undefined)
+    ?.summary?.written_rows;
+  if (written === undefined) return null;
+  const parsed = Number(written);
+  return Number.isFinite(parsed) ? parsed : null;
+}