Exam Bundle

The full-scale integration finishes without notebook OOM and matches checksums on a validated sample. This is evidence for a different corrective change: Use a suitable distributed preprocessing engine such as Spark on Dataproc for the large shuffle and persist curated outputs. That other change addresses: A data science team attempts to join terabytes of distributed logs inside a single notebook's pandas process and repeatedly exhausts memory. It does not validate the specific Silverline Insurance remediation for: A serving team scales revenue by thousands while training used unscaled currency, creating train-serving skew. The necessary check instead is: Matching test inputs generate the same feature vectors offline and online within documented numeric tolerance.

A backward-compatible new optional field passes, while removing a required field fails fast with a clear error. This is evidence for a different corrective change: Adopt compatible schema evolution with explicit validation of required fields and controlled handling of unknown optional properties. That other change addresses: A transaction feed adds optional properties weekly but the downstream feature job breaks whenever the JSON payload has a new field. It does not validate the specific Silverline Insurance remediation for: A serving team scales revenue by thousands while training used unscaled currency, creating train-serving skew. The necessary check instead is: Matching test inputs generate the same feature vectors offline and online within documented numeric tolerance.

A deliberate category-type change triggers the rejection path while schema-valid data reaches the curated table. This is evidence for a different corrective change: Validate type, required fields, and allowed categories at ingestion and quarantine incompatible batches. That other change addresses: A product classification ingestion job silently coerces the string category column into null after upstream schema changes. It does not validate the specific Silverline Insurance remediation for: A serving team scales revenue by thousands while training used unscaled currency, creating train-serving skew. The necessary check instead is: Matching test inputs generate the same feature vectors offline and online within documented numeric tolerance.

Define per-feature missingness rules and train-only imputers with explicit missing indicators where justified. This action could be suitable for a different engineering incident: An ML feed has meaningful null values but a preprocessing pipeline replaces every missing numeric measurement with zero, changing business semantics. It does not directly correct the reported Westhaven Marketplace failure: An image classification team stores training images directly in BigQuery rows as base64 strings, making batch loading inefficient and costly. The discriminating acceptance check here is: A batch loader retrieves images by reference with valid labels and no redundant large inline binary columns.

Build an event-time-aware streaming pipeline using Dataflow with watermarks, lateness handling, and idempotent writes. This action could be suitable for a different engineering incident: A mobile events platform receives late, out-of-order user events but runs a daily file-only ETL that misses required real-time features. It does not directly correct the reported Westhaven Marketplace failure: An image classification team stores training images directly in BigQuery rows as base64 strings, making batch loading inefficient and costly. The discriminating acceptance check here is: A batch loader retrieves images by reference with valid labels and no redundant large inline binary columns.

Deduplicate by stable event identity and event-version policy before constructing training features. This action could be suitable for a different engineering incident: A clickstream ingestion job double-counts purchases whenever the publisher retries a delivered event with the same event identifier. It does not directly correct the reported Westhaven Marketplace failure: An image classification team stores training images directly in BigQuery rows as base64 strings, making batch loading inefficient and costly. The discriminating acceptance check here is: A batch loader retrieves images by reference with valid labels and no redundant large inline binary columns.

The held-out data preserves real zeros as distinct from missing readings and passes predeclared data-quality checks. This is evidence for a different corrective change: Define per-feature missingness rules and train-only imputers with explicit missing indicators where justified. That other change addresses: An ML feed has meaningful null values but a preprocessing pipeline replaces every missing numeric measurement with zero, changing business semantics. It does not validate the specific Westhaven Marketplace remediation for: An image classification team stores training images directly in BigQuery rows as base64 strings, making batch loading inefficient and costly. The necessary check instead is: A batch loader retrieves images by reference with valid labels and no redundant large inline binary columns.

Replayed input produces one canonical purchase event and unchanged feature totals on repeated runs. This is evidence for a different corrective change: Deduplicate by stable event identity and event-version policy before constructing training features. That other change addresses: A clickstream ingestion job double-counts purchases whenever the publisher retries a delivered event with the same event identifier. It does not validate the specific Westhaven Marketplace remediation for: An image classification team stores training images directly in BigQuery rows as base64 strings, making batch loading inefficient and costly. The necessary check instead is: A batch loader retrieves images by reference with valid labels and no redundant large inline binary columns.

Injected out-of-order events reach the correct windows without duplicate feature records under replay. This is evidence for a different corrective change: Build an event-time-aware streaming pipeline using Dataflow with watermarks, lateness handling, and idempotent writes. That other change addresses: A mobile events platform receives late, out-of-order user events but runs a daily file-only ETL that misses required real-time features. It does not validate the specific Westhaven Marketplace remediation for: An image classification team stores training images directly in BigQuery rows as base64 strings, making batch loading inefficient and costly. The necessary check instead is: A batch loader retrieves images by reference with valid labels and no redundant large inline binary columns.

Use in-memory Python preprocessing for the bounded dataset while recording inputs and transform versions for reproducibility. This action could be suitable for a different engineering incident: A small exploratory tabular dataset is routed through a complex distributed cluster that takes longer to initialize than the analysis itself. It does not directly correct the reported Sunset Commerce failure: An image classification team stores training images directly in BigQuery rows as base64 strings, making batch loading inefficient and costly. The discriminating acceptance check here is: A batch loader retrieves images by reference with valid labels and no redundant large inline binary columns.

Validate type, required fields, and allowed categories at ingestion and quarantine incompatible batches. This action could be suitable for a different engineering incident: A product classification ingestion job silently coerces the string category column into null after upstream schema changes. It does not directly correct the reported Sunset Commerce failure: An image classification team stores training images directly in BigQuery rows as base64 strings, making batch loading inefficient and costly. The discriminating acceptance check here is: A batch loader retrieves images by reference with valid labels and no redundant large inline binary columns.

Define per-feature missingness rules and train-only imputers with explicit missing indicators where justified. This action could be suitable for a different engineering incident: An ML feed has meaningful null values but a preprocessing pipeline replaces every missing numeric measurement with zero, changing business semantics. It does not directly correct the reported Sunset Commerce failure: An image classification team stores training images directly in BigQuery rows as base64 strings, making batch loading inefficient and costly. The discriminating acceptance check here is: A batch loader retrieves images by reference with valid labels and no redundant large inline binary columns.

The held-out data preserves real zeros as distinct from missing readings and passes predeclared data-quality checks. This is evidence for a different corrective change: Define per-feature missingness rules and train-only imputers with explicit missing indicators where justified. That other change addresses: An ML feed has meaningful null values but a preprocessing pipeline replaces every missing numeric measurement with zero, changing business semantics. It does not validate the specific Sunset Commerce remediation for: An image classification team stores training images directly in BigQuery rows as base64 strings, making batch loading inefficient and costly. The necessary check instead is: A batch loader retrieves images by reference with valid labels and no redundant large inline binary columns.

The notebook processes the complete sample within memory limits and reproduces the same cleaned output without cluster startup overhead. This is evidence for a different corrective change: Use in-memory Python preprocessing for the bounded dataset while recording inputs and transform versions for reproducibility. That other change addresses: A small exploratory tabular dataset is routed through a complex distributed cluster that takes longer to initialize than the analysis itself. It does not validate the specific Sunset Commerce remediation for: An image classification team stores training images directly in BigQuery rows as base64 strings, making batch loading inefficient and costly. The necessary check instead is: A batch loader retrieves images by reference with valid labels and no redundant large inline binary columns.

A deliberate category-type change triggers the rejection path while schema-valid data reaches the curated table. This is evidence for a different corrective change: Validate type, required fields, and allowed categories at ingestion and quarantine incompatible batches. That other change addresses: A product classification ingestion job silently coerces the string category column into null after upstream schema changes. It does not validate the specific Sunset Commerce remediation for: An image classification team stores training images directly in BigQuery rows as base64 strings, making batch loading inefficient and costly. The necessary check instead is: A batch loader retrieves images by reference with valid labels and no redundant large inline binary columns.

Centralize versioned, governed feature definitions in Gemini Enterprise Agent Platform Feature Store where supported. This action could be suitable for a different engineering incident: Multiple training teams recompute inconsistent customer lifetime value features and cannot share the exact current definitions with serving. It does not directly correct the reported Harbor Rail Logistics failure: A transaction feed adds optional properties weekly but the downstream feature job breaks whenever the JSON payload has a new field. The discriminating acceptance check here is: A backward-compatible new optional field passes, while removing a required field fails fast with a clear error.

Persist the preprocessing artifact and reuse the same transform in training and inference paths. This action could be suitable for a different engineering incident: A serving team scales revenue by thousands while training used unscaled currency, creating train-serving skew. It does not directly correct the reported Harbor Rail Logistics failure: A transaction feed adds optional properties weekly but the downstream feature job breaks whenever the JSON payload has a new field. The discriminating acceptance check here is: A backward-compatible new optional field passes, while removing a required field fails fast with a clear error.

Implement point-in-time feature joins that exclude signals produced after each prediction cutoff. This action could be suitable for a different engineering incident: An account fraud model joins labels to activity from after the prediction timestamp, accidentally giving the model future knowledge. It does not directly correct the reported Harbor Rail Logistics failure: A transaction feed adds optional properties weekly but the downstream feature job breaks whenever the JSON payload has a new field. The discriminating acceptance check here is: A backward-compatible new optional field passes, while removing a required field fails fast with a clear error.

Matching test inputs generate the same feature vectors offline and online within documented numeric tolerance. This is evidence for a different corrective change: Persist the preprocessing artifact and reuse the same transform in training and inference paths. That other change addresses: A serving team scales revenue by thousands while training used unscaled currency, creating train-serving skew. It does not validate the specific Harbor Rail Logistics remediation for: A transaction feed adds optional properties weekly but the downstream feature job breaks whenever the JSON payload has a new field. The necessary check instead is: A backward-compatible new optional field passes, while removing a required field fails fast with a clear error.

A temporal audit finds zero post-cutoff feature values in training and online replay samples. This is evidence for a different corrective change: Implement point-in-time feature joins that exclude signals produced after each prediction cutoff. That other change addresses: An account fraud model joins labels to activity from after the prediction timestamp, accidentally giving the model future knowledge. It does not validate the specific Harbor Rail Logistics remediation for: A transaction feed adds optional properties weekly but the downstream feature job breaks whenever the JSON payload has a new field. The necessary check instead is: A backward-compatible new optional field passes, while removing a required field fails fast with a clear error.

Training and online reads for a controlled customer resolve the same declared feature version and transformation semantics. This is evidence for a different corrective change: Centralize versioned, governed feature definitions in Gemini Enterprise Agent Platform Feature Store where supported. That other change addresses: Multiple training teams recompute inconsistent customer lifetime value features and cannot share the exact current definitions with serving. It does not validate the specific Harbor Rail Logistics remediation for: A transaction feed adds optional properties weekly but the downstream feature job breaks whenever the JSON payload has a new field. The necessary check instead is: A backward-compatible new optional field passes, while removing a required field fails fast with a clear error.

Tokenize or pseudonymize sensitive identifiers and restrict the reversible mapping to a separately authorized service. This action could be suitable for a different engineering incident: A lending analyst copies raw national identifiers into shared training tables even though the model needs only longitudinal grouping. It does not directly correct the reported Aster Pharma failure: A transaction feed adds optional properties weekly but the downstream feature job breaks whenever the JSON payload has a new field. The discriminating acceptance check here is: A backward-compatible new optional field passes, while removing a required field fails fast with a clear error.

Deduplicate by stable event identity and event-version policy before constructing training features. This action could be suitable for a different engineering incident: A clickstream ingestion job double-counts purchases whenever the publisher retries a delivered event with the same event identifier. It does not directly correct the reported Aster Pharma failure: A transaction feed adds optional properties weekly but the downstream feature job breaks whenever the JSON payload has a new field. The discriminating acceptance check here is: A backward-compatible new optional field passes, while removing a required field fails fast with a clear error.

Partition the training table on event date and cluster on useful access keys, then apply explicit partition filters in training queries. This action could be suitable for a different engineering incident: A purchase prediction team scans a five-year BigQuery click table without date filters on every experiment and repeatedly incurs avoidable full-table reads. It does not directly correct the reported Aster Pharma failure: A transaction feed adds optional properties weekly but the downstream feature job breaks whenever the JSON payload has a new field. The discriminating acceptance check here is: A backward-compatible new optional field passes, while removing a required field fails fast with a clear error.

Query plans show partition pruning and reduced scanned bytes while experiment outputs retain the expected record counts. This is evidence for a different corrective change: Partition the training table on event date and cluster on useful access keys, then apply explicit partition filters in training queries. That other change addresses: A purchase prediction team scans a five-year BigQuery click table without date filters on every experiment and repeatedly incurs avoidable full-table reads. It does not validate the specific Aster Pharma remediation for: A transaction feed adds optional properties weekly but the downstream feature job breaks whenever the JSON payload has a new field. The necessary check instead is: A backward-compatible new optional field passes, while removing a required field fails fast with a clear error.

Replayed input produces one canonical purchase event and unchanged feature totals on repeated runs. This is evidence for a different corrective change: Deduplicate by stable event identity and event-version policy before constructing training features. That other change addresses: A clickstream ingestion job double-counts purchases whenever the publisher retries a delivered event with the same event identifier. It does not validate the specific Aster Pharma remediation for: A transaction feed adds optional properties weekly but the downstream feature job breaks whenever the JSON payload has a new field. The necessary check instead is: A backward-compatible new optional field passes, while removing a required field fails fast with a clear error.

Training records carry nonreversible working tokens and permission tests prevent analysts from reading raw identifiers. This is evidence for a different corrective change: Tokenize or pseudonymize sensitive identifiers and restrict the reversible mapping to a separately authorized service. That other change addresses: A lending analyst copies raw national identifiers into shared training tables even though the model needs only longitudinal grouping. It does not validate the specific Aster Pharma remediation for: A transaction feed adds optional properties weekly but the downstream feature job breaks whenever the JSON payload has a new field. The necessary check instead is: A backward-compatible new optional field passes, while removing a required field fails fast with a clear error.

Implement point-in-time feature joins that exclude signals produced after each prediction cutoff. This action could be suitable for a different engineering incident: An account fraud model joins labels to activity from after the prediction timestamp, accidentally giving the model future knowledge. It does not directly correct the reported Mercury Travel Tech failure: A clickstream ingestion job double-counts purchases whenever the publisher retries a delivered event with the same event identifier. The discriminating acceptance check here is: Replayed input produces one canonical purchase event and unchanged feature totals on repeated runs.

Use event-time or point-in-time feature retrieval against effective-dated feature values. This action could be suitable for a different engineering incident: A recommendation model joins current customer status to old events, creating implausible historical feature values from updated records. It does not directly correct the reported Mercury Travel Tech failure: A clickstream ingestion job double-counts purchases whenever the publisher retries a delivered event with the same event identifier. The discriminating acceptance check here is: Replayed input produces one canonical purchase event and unchanged feature totals on repeated runs.

Tokenize or pseudonymize sensitive identifiers and restrict the reversible mapping to a separately authorized service. This action could be suitable for a different engineering incident: A lending analyst copies raw national identifiers into shared training tables even though the model needs only longitudinal grouping. It does not directly correct the reported Mercury Travel Tech failure: A clickstream ingestion job double-counts purchases whenever the publisher retries a delivered event with the same event identifier. The discriminating acceptance check here is: Replayed input produces one canonical purchase event and unchanged feature totals on repeated runs.

Training records carry nonreversible working tokens and permission tests prevent analysts from reading raw identifiers. This is evidence for a different corrective change: Tokenize or pseudonymize sensitive identifiers and restrict the reversible mapping to a separately authorized service. That other change addresses: A lending analyst copies raw national identifiers into shared training tables even though the model needs only longitudinal grouping. It does not validate the specific Mercury Travel Tech remediation for: A clickstream ingestion job double-counts purchases whenever the publisher retries a delivered event with the same event identifier. The necessary check instead is: Replayed input produces one canonical purchase event and unchanged feature totals on repeated runs.

Historical replay retrieves status as it was at each event time, rather than the latest account record. This is evidence for a different corrective change: Use event-time or point-in-time feature retrieval against effective-dated feature values. That other change addresses: A recommendation model joins current customer status to old events, creating implausible historical feature values from updated records. It does not validate the specific Mercury Travel Tech remediation for: A clickstream ingestion job double-counts purchases whenever the publisher retries a delivered event with the same event identifier. The necessary check instead is: Replayed input produces one canonical purchase event and unchanged feature totals on repeated runs.

A temporal audit finds zero post-cutoff feature values in training and online replay samples. This is evidence for a different corrective change: Implement point-in-time feature joins that exclude signals produced after each prediction cutoff. That other change addresses: An account fraud model joins labels to activity from after the prediction timestamp, accidentally giving the model future knowledge. It does not validate the specific Mercury Travel Tech remediation for: A clickstream ingestion job double-counts purchases whenever the publisher retries a delivered event with the same event identifier. The necessary check instead is: Replayed input produces one canonical purchase event and unchanged feature totals on repeated runs.