What Prizm Collects from Snowflake

Prizm runs the following platform jobs against each Snowflake connector. Each job has a defined type, execution flow, and purpose. Select a job below to see the details.


Type	CONTEXT
Execution	Runs immediately after source is configured — first job in the pipeline

What this job does: Extracts accounts, warehouses, and databases from Snowflake. This is the foundation job — all subsequent jobs depend on it completing successfully. Without a successful TECHNICAL run, no other job can start.

Object	Fields Collected	Snowflake Source
Account	Account name, region, cloud provider	`SHOW ACCOUNTS`
Warehouse	Name, size, type, state, created_on	`SHOW WAREHOUSES`
Database	Name, created_on, owner, comment	`SHOW DATABASES`
Schema	Name, database, created_on, owner	`INFORMATION_SCHEMA.SCHEMATA`
Table	Name, type, last_altered, row_count, is_iceberg, is_hybrid	`INFORMATION_SCHEMA.TABLES`
View	Same as Table + view_definition	`INFORMATION_SCHEMA.VIEWS`
Column	Name, data_type, ordinal_position, is_nullable	`INFORMATION_SCHEMA.COLUMNS`

Prizm catalogs TABLE and VIEW object types. Dynamic Tables, Stored Procedures, External Tables, and Materialized Views are not currently included in the catalog collection.


Type	CONTEXT / TRUST
Execution	Starts after TECHNICAL completes

What this job does: Extracts schemas, tables, views, and volume, freshness, and schema information for table assets from Snowflake metadata. Serves a dual purpose — it feeds the Prizm catalog (CONTEXT) and the alerting and anomaly detection pipeline (TRUST). It is the source of all time-series observability signals in Prizm and the upstream dependency for most parallel jobs.

Signal	Source	Description
Volume	`INFORMATION_SCHEMA.TABLES` (row_count)	Row count trend; anomaly alerts on drops/spikes
Freshness	`INFORMATION_SCHEMA.TABLES` (last_altered)	Time since last DDL or DML update; SLA breach alerts
Schema Change	`INFORMATION_SCHEMA.TABLES` (last_ddl)	Column additions, renames, type changes, removals

Anomaly detection uses adaptive forecasting models (EWMA, AutoReg, Prophet, SARIMAX) selected automatically based on each metric’s historical pattern. Thresholds are recalculated after every run. A new metric spends its first more than 5 historical collection runs in Learning mode before alerts begin firing — this threshold is counted in collection runs, not elapsed time.

Pausing and resuming a source. Freshness is computed at collection time as the interval since the asset was last altered. Pausing collection freezes the displayed freshness value at whatever it was when the pause took effect — Prizm has no way to know the asset went untouched during the pause versus simply not being checked. When collection resumes, the first run measures the entire pause duration as staleness, which typically fires a critical freshness alert on every asset at once and writes that spike into the anomaly-detection baselines, distorting future thresholds.To avoid both the false alert storm and the baseline contamination when resuming a paused source: let the source’s underlying tables receive normal activity first, then let one full collection cycle complete and land before re-enabling downstream alerting, rather than re-enabling everything the moment the source itself reconnects.


Type	CONTEXT
Execution	No scheduled run. Lineage does not appear on the schedule screen and does not fire automatically — it runs only when triggered manually from an individual asset’s run menu.

What this job does: Extracts lineage using Snowflake ACCESS_HISTORY and QUERY_HISTORY to identify object dependencies and data flow.

Lineage has no automatic schedule today. Trigger it manually per asset when you need current lineage data. Snowflake’s underlying OBJECT_DEPENDENCIES view can lag up to 3 hours behind the actual DDL change, so a manual run immediately after a schema change may not yet reflect it — wait for the latency window to pass before re-running if the lineage graph looks stale.

Run	Method
First run — view lineage	Seeded from `SNOWFLAKE.ACCOUNT_USAGE.OBJECT_DEPENDENCIES`
First run — table lineage	Extracted per asset using `SNOWFLAKE.CORE.GET_LINEAGE`
Subsequent runs (DDL change detected)	Re-extracted only for assets with DDL changes; column-level lineage included (`include_attribute_lineage = true`)

Supported asset types: TABLE, VIEW, EXTERNAL_TABLE, MATERIALIZED_VIEW. Default batch size 100, concurrency 8.

Direction	What Prizm Shows
Upstream	Tables and views this asset reads from
Downstream	Tables, views, and pipelines that depend on this asset

Lineage — both table-level and column-level — requires Snowflake Enterprise edition or above, since it depends on SNOWFLAKE.CORE.GET_LINEAGE. Standard edition does not have a query-log-parsing fallback: on a Standard-edition account, the Lineage graph stays empty.


Type	CONTEXT
Execution	Runs in parallel after OPERATIONAL completes

What this job does: Extracts warehouse usage, credit consumption, and query statistics for the previous 24 hours. Powers the Usage and Performance views in the Prizm asset detail page. Requires IMPORTED PRIVILEGES on the SNOWFLAKE database.

Category	Metrics	Source
Query Performance	Execution time, queue time, compilation time, bytes scanned, bytes written, rows produced, status	`ACCOUNT_USAGE.QUERY_HISTORY`
Warehouse Compute	Credits used per warehouse per hour	`ACCOUNT_USAGE.WAREHOUSE_METERING_HISTORY`
Storage	Average bytes stored (tables, fail-safe, stage)	`ACCOUNT_USAGE.DATABASE_STORAGE_USAGE_HISTORY`
Data Transfer	Bytes transferred to external destinations	`ACCOUNT_USAGE.DATA_TRANSFER_HISTORY`
Access History	User, query, objects read/written, timestamp	`ACCOUNT_USAGE.ACCESS_HISTORY`

ACCOUNT_USAGE views have a data latency of up to 45 minutes. Prizm retains time-series observability data, quality score history, and query usage history for a default window of 90 days.


Type	CONTEXT
Execution	Runs in parallel after OPERATIONAL completes

What this job does: Extracts summarized usage information and generates usage-based criticality inputs using 7 days of activity. Analyses how frequently each asset is queried, by whom, and with what patterns — then feeds those signals into the CRITICALITY RECOMMENDATION job. Assets with high query volume or many unique users are more likely to be recommended as critical.

What It Produces	Description
Usage frequency	Number of queries per asset over the past 7 days
Unique user count	Number of distinct users who accessed each asset
Access patterns	Read vs write activity, peak usage windows
Criticality inputs	Usage-based signals consumed by the CRITICALITY RECOMMENDATION job


Type	CONTEXT
Execution	Runs in parallel after OPERATIONAL completes

What this job does: Extracts governance metadata including tags and business metadata. Keeps Prizm’s governance layer in sync with classifications, ownership, and tagging applied directly in Snowflake. Requires IMPORTED PRIVILEGES on the SNOWFLAKE shared database for tag ingestion.

Object	Fields Collected	Snowflake Source
Tags(Enterprise)	Tag name, value, associated objects	`ACCOUNT_USAGE.TAGS`, `TAG_REFERENCES`
Business metadata	Owner, description, classification fields	`INFORMATION_SCHEMA` object properties


Type	CONTEXT
Execution	Runs in parallel after OPERATIONAL completes

What this job does: Calculates and recommends criticality scores for assets and attributes. Criticality determines how much observability coverage an asset receives — higher criticality assets get more frequent monitoring, more structural metrics, and tighter alert thresholds. The output is a recommended criticality level (LOW, MEDIUM, HIGH, CRITICAL) per asset, which users can accept or override.

Input Signal	Source Job
Usage frequency and user count	USAGE SUMMARY
Downstream dependency count	LINEAGE
Business metadata completeness	GOVERNANCE
Asset type and schema position	TECHNICAL / OPERATIONAL


Type	TRUST
Execution	Runs in parallel after OPERATIONAL completes

What this job does: Computes non-metadata metrics such as freshness and volume for non-table assets by running direct queries against Snowflake. The OPERATIONAL job derives these signals from metadata (e.g. INFORMATION_SCHEMA.TABLES.LAST_ALTERED) — but for assets like views and external tables, metadata alone is not sufficient. COMPUTE METRIC fills this gap by executing lightweight SQL queries to obtain accurate signals.

Asset Type	What Is Computed	How
View	Row count, freshness	Direct `SELECT COUNT(*)` and query history lookup
External Table	Row count, last modified	Query against external stage metadata
Materialized View	Row count, refresh timestamp	Query against materialized view metadata


Type	TRUST
Execution	Runs in parallel after OPERATIONAL; PROFILE job depends on its output

What this job does: Recommends profiling schedules and configurations based on asset scope definitions. Determines which assets should be profiled, at what frequency, and with what sample settings — taking into account criticality level, asset size, and scope configuration. The PROFILE job will not run until this job has produced its output and created the schedules.


Type	TRUST
Execution	Runs in parallel after OPERATIONAL; STRUCTURAL job depends on its output

What this job does: Recommends structural quality metrics (Essential or Advanced) for each attribute. Analyses the column’s data type, profile history, and criticality to determine the appropriate metric set. Essential metrics are always recommended; Advanced metrics are added for columns where deeper analysis is warranted. The STRUCTURAL job executes against this recommendation output.


Type	TRUST
Execution	Triggered after PROFILE RECOMMENDATION completes and schedules are created

What this job does: Executes profiling jobs according to the schedules generated by PROFILE RECOMMENDATION and stores the profiling results. Can also be triggered manually from the asset page at any time.Table-level metrics:

Metric	Description
Row Count	Total rows at profile time
Duplicate Row Count	Rows that are exact duplicates of another row
Completeness Score	% of columns with non-null values across all rows

Column-level metrics:

Metric	Applicable Types	Description
Null Rate	All	% of null values
Distinct Count	All	Number of unique non-null values (cardinality)
Uniqueness Rate	All	% of values appearing exactly once
Min / Max	Numeric, Date/Timestamp	Observed range of values
Mean	Numeric	Arithmetic mean
Standard Deviation	Numeric	Statistical spread
Median	Numeric	50th percentile value
Top N Values	String, Boolean	Most frequent values and their counts
Pattern Conformance	String	% matching an expected regex format
Data Type Conformance	String	% of values parseable as the declared type

Advanced numeric statistics are computed for types containing int, number, decimal, float, or double. Snowflake types such as REAL, BIGINT, and SMALLINT still receive min, max, distinct count, and null counts.

Trigger conditions:

Trigger	Conditions
Scheduled run	Active schedule, `next_run_at` is due, source is valid and active
Manual run	User triggers Profile from the asset UI (TABLE, VIEW, MATERIALIZED_VIEW)
Platform mode	PROFILE job type must be enabled in platform mode
Criticality gating	Only attributes eligible for the asset’s criticality level are profiled
Skip (incremental)	Skipped if row count is unchanged since last run, or profile is newer than `last_altered_on`

Supported asset types: TABLE, VIEW, MATERIALIZED_VIEW (and ATTRIBUTE for column-level context).


Type	TRUST
Execution	Triggered after STRUCTURAL RECOMMENDATION completes and metrics are created

What this job does: Executes structural quality metrics according to the schedules and configurations generated by STRUCTURAL RECOMMENDATION. Computes distribution, frequency, pattern, and statistics metrics at the attribute level for all in-scope columns.

Metric Category	What Is Computed
Distribution	Null count, empty, space, distinct, repeating, character composition, leading/trailing/inner space
Frequency	Enum (value histogram), min/max value, value range, min/max length, length distribution
Pattern	Short pattern, long pattern, user-defined regular expressions
Statistics	Mean, median, mode, standard deviation, sum, Q1, Q3, range, skewness, kurtosis, variance, margin of error


Type	CONTEXT
Execution	Independent job; no upstream dependency

What this job does: Recommends asset definitions, descriptions, intended audience, semantic context, and execution scope settings (Incremental, Sampling with Filter, or None). Uses AI to analyse the asset’s name, column names, sample values, and usage patterns to generate human-readable context that data stewards can review and accept.

What It Generates	Description
Description	AI-generated plain-language description of what the asset contains and its purpose
Who should use	Recommended audience — data analysts, engineers, business users, etc.
Semantic context	Suggested domain, product, application, and tag classifications
Execution scope	Recommended profiling mode — Incremental (only changed rows), Sampling with Filter, or None


Type	TRUST
Execution	Independent job; no upstream dependency

What this job does: Recommends business metrics for assets based on column names, data types, profiling results, and industry patterns. Suggests quality checks that go beyond structural profiling — for example, a not-null check on a primary key column, or a range check on a revenue column. Recommendations appear in the Prizm UI for review before any metric is created.


Type	TRUST
Execution	Independent job; no upstream dependency

What this job does: Recommends metric descriptions, purposes, and rationale for existing metrics. For each metric already defined on an asset, this job generates a human-readable explanation of what the metric checks, why it matters, and what a failure would indicate. These descriptions appear in the Prizm metric detail panel and help data stewards understand the intent behind auto-generated checks.


Type	CONTEXT
Execution	Independent job; no upstream dependency

What this job does: Maps assets and attributes to existing glossary terms or generates new business terms where required. Analyses column names, descriptions, and sample values to find matching terms in the Prizm business glossary. Powers the Semantic layer in Prizm — ensuring that columns are aligned to approved business vocabulary.

Output	Description
Term mapping	Links a column to an existing glossary term (e.g. `customer_id` → “Customer Identifier”)
Term proposal	Suggests a new glossary term where no existing match is found
Confidence score	Indicates how confident the AI is in the mapping recommendation


Type	TRUST
Execution	Independent job; no upstream dependency

What this job does: Clusters and groups related alerts to reduce duplication and improve root-cause analysis. When multiple alerts fire simultaneously on the same asset or related assets — for example, a freshness alert and a volume alert on the same table — this job groups them into a single alert cluster. Reduces alert fatigue by surfacing correlated failures together rather than as separate notifications.


Type	TRUST
Execution	Independent job; no upstream dependency

What this job does: Generates and refreshes daily asset trend data by aggregating DQ scores, alerts, issues, and asset metadata. Powers the score history charts on the asset detail page — the time-series view of how an asset’s quality score has evolved over time. Also feeds domain-level and portfolio-level trend views in the Prizm Scorecard and Analytics pages.


Type	PLATFORM
Execution	Independent job; no upstream dependency

What this job does: Cleans and maintains the Prizm repository by removing obsolete data. Deletes stale metric history, expired alert records, orphaned job artifacts, and data from deleted or disconnected sources. Runs on a configurable schedule to keep the Prizm database within the configured retention window. Does not affect live assets or active metric data within the retention period.

None of the four recommendation jobs above (Asset Detail, Business Metric, Metric Detail, Term) can be generated in bulk — each produces suggestions for one asset at a time. The only bulk control available is Approve Recommendations on the asset page, which accepts already-generated suggestions for that asset’s pending recommendations; it does not create new ones. There is currently no control that generates recommendations across more than one asset at a time.

What You See in Prizm

Once Prizm completes its initial catalog and observability runs, every Snowflake table and view gets a unified asset detail page. The tabs below cover what each section shows.

Overview
Attributes & Profile
Metrics
Usage
Lineage

Section	What Prizm Shows
Quality Score	Overall data quality percentage computed across all active metrics
Status	Review state (e.g., READY FOR REVIEW) and criticality badge
Key Metrics	VOLUME (row count), SCHEMA (column count), FRESHNESS (time since last update), METRICS (total quality metrics defined)
Description	AI-generated or manually verified description of the asset
Semantic Context	Domain, Application, Product, and Tag classifications
Owners	Business owner, Technical owner, and Steward contacts
Lineage Summary	One-hop upstream sources and downstream consumers
Context Panel	AI-generated narrative, context completeness score (0–100%), and freshness timestamps for description, terms, and owner verification

The Attribute tab lists every column Prizm has cataloged — data type, quality score, alert/issue counts, and the number of quality metrics defined.

Clicking any attribute opens its column profile with four views: Distribution, Pattern, Custom Pattern, and Frequency.

Profile Section	What Prizm Shows
Completeness	Count and % of NULL, EMPTY, SPACE, and NON-EMPTY values
Uniqueness	Count and % of DISTINCT vs REPEATING values
Character	Distribution across DIGIT, ALPHABET, ALPHA NUMERIC, SPECIAL CHARACTER
Length	Minimum and maximum character length
Pattern	Top regex patterns detected across values
Frequency	Most and least frequent individual values

For the full breakdown of profiling views, metric types, and how to interpret profile results, see the Data Profiling section.

The Metric tab lists all quality metrics defined on the asset, grouped by context. Each row shows the metric name, column, domain, current score, alert count, issue count, and action controls.

Asset-level health signals. Monitor: Yes — Score: No

Metric	Subcategory	Description
Row	Volume	Total number of rows at the time of last scan
Volume	Volume	Combined row count and data size signal
Data size	Volume	Physical size of the asset in bytes
Freshness	Freshness	Time elapsed since the asset was last updated
Last updated	Freshness	Timestamp of the most recent data load or write
Column	Schema	Current number of columns
Schema	Schema	Full column snapshot — names, types, and order
Schema name	Schema	Name of the schema the asset belongs to
Duplicates	Duplicate	Count of duplicate rows based on a defined key or all columns

Auto-generated column-level metrics. Score: YesDistribution

Metric	Subcategory
Null count	Completeness
Empty	Completeness
Non-empty	Completeness
Space	Completeness
Zero value	Numeric
Distinct	Uniqueness
Repeating	Uniqueness
Leading space	Space
Trailing space	Space
Inner space	Space
Outer space	Space
Whitespace	Space
Character	Character
Alpha numeric	Character
Digit	Character
Special character	Character
Alphabet	Character
Positive	Numeric
Negative	Numeric

Frequency

Metric	Subcategory
Enum	Value
Min value	Value
Max value	Value
Value range	Range
Min length	Length
Max length	Length
Length	Length
Length range	Range

Pattern

Metric	Subcategory
Short pattern	Pattern
Long pattern	Pattern
Regular expressions	Pattern

Statistics

Metric	Subcategory
Mean	Descriptive
Median	Descriptive
Mode	Descriptive
Standard deviation	Descriptive
Sum	Descriptive
Q1	Quantile
Q3	Quantile
Range	Shape
Skewness	Shape
Kurtosis	Shape
Variance	Shape
Margin of error	Shape

User-defined metrics. Custom, Conditional, Standalone: Score Yes — Behavioral: Score No

Metric Type	Description
Custom	User-defined rules, filters, or expressions on a specific asset
Conditional	Different logic or thresholds based on a condition
Query	SQL query returning a scalar — enables multi-table logic
Standalone	Not bound to a specific asset — used for org-wide KPIs
Behavioral	Tracks patterns over time — detects drift, not point-in-time violations

Platform-level signals. Monitor: Yes — Score: No

Metric	Level
Total queries per day	Warehouse, Database
Query success rate	Warehouse, Database
Execution time	Warehouse, Database
Job duration	Warehouse, Database
Test execution time	Warehouse, Database
Avg blocked	Warehouse, Database
Avg queue load	Warehouse, Database
Avg running	Warehouse, Database
Database storage	Database
Total credits per day	Warehouse, Database
Credits used	Warehouse, Database

Cross-source comparison metrics.

Metric Type	Description
Comparison	Compares values, row counts, or aggregates across two registered assets
Lookup	Validates column values against a reference table or allowed set

Domain and grouping level metrics.

Grouping	Description
Domain	DQ score rolled up across all assets assigned to a domain
Product	Aggregate quality and freshness across a defined data product
Application	% of assets in an application meeting quality and freshness SLOs
Tag	PII tag coverage across assets with sensitive columns
Connection	Aggregate quality score across all assets from a given Snowflake connection

The Usage tab surfaces per-asset query intelligence derived from the Performance job.

Metric	Description
Total Queries	Total number of queries that touched this asset
Average Execution Time	Mean query execution time in milliseconds
Query Success Rate	Ratio of successful queries to total queries
Total Consumed Credit	Snowflake credits consumed by queries against this asset
Unique Users	Number of distinct users who queried this asset
Unique Warehouses	Number of distinct warehouses used to query this asset

Queries can be filtered by EXPENSIVE, SLOWER, and POPULAR.

Next Steps

Setup

Review prerequisites and configure the connector.

FAQ

Common questions about what Prizm collects and why.

Get Started

Core Concepts

Platform

Architecture

Data Sources

Integrations

Security & Access Control

Deployment

Help

Settings

MCP

Release Notes

What Prizm Collects from Snowflake

What You See in Prizm

Next Steps

Setup

FAQ

​What You See in Prizm

​Next Steps

Setup

FAQ

What You See in Prizm

Next Steps