What you will be able to do
- Trace upstream and downstream lineage of tables, feature views, datasets and models in Snowsight and with SNOWFLAKE.CORE.GET_LINEAGE
- Grant the privileges a role needs to view lineage
- Choose Snowpipe Streaming and the right channel type for near real-time feature data
- Attach system and custom DMFs with expectations and set the schedule they run on
- Use the Feature Store monitoring view to track freshness, refresh failures, ingest errors and schema mismatches, and know where drift checks have to go instead
Key concept
Data quality check (DMF + expectation) — A data metric function only measures something about your data, such as a NULL count or how recently a table was updated. It becomes a pass/fail check only when you pair it with an expectation, a Boolean rule that each returned value is compared against.
1.Tracing lineage from raw tables to models
Before you change a raw table that feeds features, you need to know what sits downstream of it. Snowflake records data lineage, meaning how data flows from source objects to target objects, so you can answer that before the change instead of after a broken refresh. Lineage records two kinds of relationship. Data movement is when data is copied or materialized from one object into another, as with CREATE TABLE AS SELECT, INSERT or MERGE. Object dependencies are when an object references a base object without copying it, as a view references a table. A training dataset or model built by copying data is a data-movement edge. A tool that only lists references between objects will miss it, and lineage will not.
In Snowsight, open Catalog » Explorer, select a table or view, and switch to the Lineage tab. The graph shows one step upstream or downstream at a time, and the +/- controls expand it further. Select an arrow between two objects to see how the downstream object was created, for example the SQL statement behind it, or the direct and root stored procedure if a procedure or task produced it. Two gaps are worth knowing. A stored procedure called anonymously leaves no procedure details in the lineage. Procedure and task details are also not backfilled for lineage recorded before that support existed.
Column lineage is how you size up a column rename. Select the object, hover over the column in the side panel, choose View Lineage, then pick Downstream Lineage. Every column that shares lineage is listed with a Distance value. A distance of 1 means the object was created directly from this one, and 2 means it was created from an object that was itself created from this one. The same dialog flags tags that are missing or mismatched on upstream and downstream columns and lets you apply them with Review and Apply.
To get the same graph as rows, call the table function SNOWFLAKE.CORE.GET_LINEAGE. It takes three required arguments: the object name (fully qualified if it is outside the current schema), its domain, and a direction of 'UPSTREAM' or 'DOWNSTREAM'. Feature views use the 'TABLE' domain, while datasets and models use 'DATASET' and 'MODEL'. For a versioned object, such as a dataset or model, you can pass a version, and the default version is used if you don't. max_distance sets how many levels to return. Both its maximum and its default are 5.
SNOWFLAKE.CORE.GET_LINEAGE(
'<object_name>',
'<object_domain>',
'<direction>',
[ <max_distance>, ]
[ '<object_version>', ]
[ '<namespace>', ]
[ '<object_type>', ]
[ '<external_id>', ]
[ '<column_name>' ]
)Each output row is one edge, a source object and a target object. Each row has domains such as TABLE, DATASET or MODEL, a version column for versioned objects, a DISTANCE, and a PROCESS variant that can hold the query ID or stored procedure that moved the data. An empty result means no lineage was recorded, and it is not an error. Errors come from an object that does not exist, is not accessible, does not support lineage, or is in a different domain than the one you passed. Output stops at 10 million edges, and anything beyond that is silently truncated.
Checkpoint 1 of 9· Fill the gap
You want every object built from a raw orders table, returned as one row per edge in the lineage graph. Which name completes the call?
SNOWFLAKE.CORE. ? (
'<object_name>',
'<object_domain>',
'<direction>',
[ <max_distance>, ]
[ '<object_version>', ]
[ '<namespace>', ]
[ '<object_type>', ]
[ '<external_id>', ]
[ '<column_name>' ]
)GET_LINEAGE is the table function that returns lineage upstream or downstream of an object. It has three required arguments: the object name, its domain, and a direction of 'UPSTREAM' or 'DOWNSTREAM'. If you leave out the optional ones, max_distance defaults to 5 and the default version is used.
Source: docs.snowflake.comCheckpoint 2 of 9· Exam question
A data engineer plans to rename the amount_usd column in a raw orders table that feeds several feature views, datasets, and registered models. Which TWO approaches identify the downstream ML objects that the change would affect?(Select 2)
Correct answers: A, E — Open the table in Snowsight, switch to the Lineage tab, and trace column lineage for amount_usd downstream to feature views, datasets, and models.; Call SNOWFLAKE.CORE.GET_LINEAGE with the table's qualified name, its object domain, and the DOWNSTREAM direction to list dependent objects.
- A. Correct: Snowsight lineage covers tables, feature views, datasets, and models, and column lineage shows which downstream columns derive from the renamed one.
- B. Incorrect: that view records definitional dependencies such as a view referencing a table, and it does not track datasets or models built from data copied out of the table.
- C. Incorrect: grants say who may read the table, not which objects were derived from it, so roles do not map to downstream features or models.
- D. Incorrect: text search over a short window misses consumers that ran outside it and cannot link queries to registered models or datasets.
- E. Correct: the GET_LINEAGE table function exposes the same lineage graph programmatically, so downstream feature views, datasets, and models can be listed in a query.
2.Privileges that let a role see lineage
Viewing lineage takes three privileges together. A role needs VIEW LINEAGE on the account. It needs some privilege on each object whose lineage it evaluates, such as SELECT on a table. It also needs USAGE or another privilege on the containing database and schema. If someone should see an object's lineage without reading its data, grant REFERENCES on the object instead of SELECT.
By default the PUBLIC role holds VIEW LINEAGE, so everyone can view lineage. To narrow that, revoke the privilege from PUBLIC and grant it to custom roles. For a governance role that has to see the whole graph, including objects it has no privileges on, grant RESOLVE ALL on the account. That role still needs VIEW LINEAGE. Where a viewer lacks access to an upstream or downstream object, the node appears gray, and Snowsight places such objects in a separate group that is collapsed by default.
Checkpoint 3 of 9· Check yourself
An analyst's role can read a raw orders table but has no privileges on one of the feature views built from it. The analyst opens the table's Lineage tab in Snowsight. How does that feature view show up?
Lineage still shows an object the viewer can't access, but only in limited form. Snowsight puts it in its own group of inaccessible objects and collapses that group by default. It doesn't drop the object, fail the whole graph, or reveal the object's details.
“Objects you don’t have access to are placed in a separate group that is collapsed by default.”Source: docs.snowflake.com
Checkpoint 4 of 9· Exam question
Which privilege setup lets a user view data lineage for objects in Snowsight?
Correct answer: D — VIEW LINEAGE on the account, object-level privileges such as SELECT or REFERENCES on the objects, and USAGE on the containing database and schema.
- A. Incorrect: lineage is captured by the platform automatically and is not computed by user-owned tasks or a warehouse.
- B. Incorrect: that grant exposes shared usage views but does not confer VIEW LINEAGE, so the Lineage graph still withholds nodes.
- C. Incorrect: MONITOR does not control lineage, and ownership of every node is not required; read-level privileges on the objects suffice.
- D. Correct: lineage is gated by the account-level VIEW LINEAGE privilege and the user also needs access to the objects shown, including USAGE on their database and schema.
Sources1
3.Snowpipe Streaming for near real-time features
When events arrive one row at a time from an application, and a feature has to reflect them within seconds, the ingestion path decides the latency. Snowpipe Streaming lets producers stream rows from devices, applications and services straight into Snowflake tables or Snowflake-managed Iceberg tables. Rows can become queryable in as little as 5 seconds, and throughput reaches up to 20 GB/s per table. No staging files or intermediate object storage are involved. The service is serverless, scales with ingestion load, and is billed per uncompressed GB ingested.
Classic Snowpipe solves a different problem. It fits a pipeline that already produces files in cloud storage, where batch-oriented, higher-latency loading is acceptable. Making an application write one file per event, just to trigger file-based loading, adds the staging hop that Snowpipe Streaming exists to remove.
Every row travels through a channel, a logical path through a PIPE object to the target table. Choose the channel type from what the source needs:
| Channel type | Delivery | Who manages channels | Use when |
|---|---|---|---|
| Elastic Channels | At-least-once, no ordering guarantee | Snowflake manages them and scales with traffic | Most new applications, IoT and telemetry producers writing directly |
| Named Channels | Ordered, exactly-once within each channel via offset tokens | The application, mapped to source partitions | Sources needing strict ordering, such as Kafka partitions or CDC |
Producers write with the Java SDK (Java 11+), the Python SDK (Python 3.9+), the Node.js SDK (Node.js 20+), the REST API, or the Snowflake Connector for Kafka. Prefer an SDK where you can. The SDKs buffer and batch appends using time and size thresholds and handle compression, while REST clients have to build NDJSON payloads and compress them themselves. The PIPE object can also transform rows in flight with COPY command syntax, reordering columns, casting types and applying expressions before commit, so light cleansing needs no separate ETL step. With schema evolution, new columns that appear in the stream are added to the table without manual DDL. A new column can therefore show up downstream without anyone running ALTER TABLE.
The streaming table is usually the source for a feature computed on a schedule. A dynamic table over it refreshes to its declared target lag without any orchestration you build (Dynamic table refresh policies and target lag). Ingestion progress and errors are visible by querying event table telemetry, and with error logging enabled you can inspect or reprocess rejected rows.
Checkpoint 5 of 9· Check yourself
A checkout service emits one event per purchase, and a fraud feature has to see each purchase within seconds. The team currently has no files landing in cloud storage. Which ingestion design fits?
The data arrives as rows from an application and has to be available with low latency, which is the case Snowpipe Streaming is built for. Classic Snowpipe is for pipelines that already produce files and can accept batch-oriented, higher-latency loading. Writing one file per event only adds the staging hop that Snowpipe Streaming removes.
“Use Snowpipe Streaming when data arrives as rows from applications, devices, or services and you need low-latency data availability.”Source: docs.snowflake.com
Checkpoint 6 of 9· Exam question
A fraud team receives card-swipe events from an application as individual rows and needs them queryable in a feature table within seconds, without staging files in cloud storage. Which ingestion approach fits?
Correct answer: A — Snowpipe Streaming, writing rows with a client SDK straight into a PIPE-backed table so no intermediate files are staged.
- A. Correct: Snowpipe Streaming is built for row-oriented, low-latency ingestion directly from applications, with data queryable in seconds.
- B. Incorrect: a COPY task is batch and file-based, and the one-minute cadence cannot deliver the seconds-level freshness required.
- C. Incorrect: classic Snowpipe is file-based and adds staging and notification latency; per-event files would also be inefficient.
- D. Incorrect: per-event inserts keep a warehouse running and do not scale; Snowflake provides a purpose-built streaming path instead.
Sources3
4.Validating ingestion tables with DMFs and expectations
Rows that land in seconds can be wrong just as fast, so quality checks run on the ingestion table itself. A data metric function (DMF) measures one attribute, such as the NULL count in a column or how often the table is updated. Snowflake ships system DMFs for common metrics in the SNOWFLAKE.CORE schema. A DMF only returns a number, and it does not decide whether that number is a problem. That decision comes from an expectation, a Boolean expression over the keyword VALUE. When the expression is TRUE the check passed, and failing values are reported as expectation violations.
ALTER VIEW v1
ADD DATA METRIC FUNCTION SNOWFLAKE.CORE.NULL_COUNT ON (C1)
EXPECTATION my_exp ( VALUE < 10);How often checks run is set by the DATA_METRIC_SCHEDULE parameter on the table or view, and it defaults to once an hour. You can set it as a number of minutes or as a cron expression. You can also make it a trigger on DML changes, such as inserted rows, though only for certain kinds of tables, and reclustering does not count as a trigger. Setting the schedule to an empty string suspends the object's DMFs, and UNSET restores the 60-minute default.
ALTER TABLE hr.tables.empl_info SET DATA_METRIC_SCHEDULE = '5 MINUTE';Scheduled DMFs run on serverless compute and appear on the bill as Data Quality Monitoring. Creating a DMF is free, and so is calling one ad hoc in a SELECT. Only scheduled computation is billed. DMFs can be set on tables (including temporary and transient), views, materialized views, dynamic tables, event tables, external tables and Iceberg tables, but not on hybrid tables or streams. An account can hold 50,000 DMF associations in total. DMFs cannot be set on shared tables or in reader accounts, and trial accounts don't support the feature.
Checkpoint 7 of 9· Check yourself
On a streaming ingestion table, a team attaches SNOWFLAKE.CORE.NULL_COUNT to the customer_id column without an EXPECTATION clause. The next scheduled run returns 250. What does Snowflake do with that result?
A DMF only measures. Pass or fail comes from an expectation, a Boolean expression over VALUE, and only values that fail an expectation are reported as violations. Without an expectation, 250 is just a number.
“The DMF returns a value based on the current state of your data, but doesn’t define whether that value constitutes a data quality issue”Source: docs.snowflake.com
5.Feature validation rules beyond basic checks
Feature rules are often specific to the business, for example that all three amount columns are positive, or that every transaction's card exists in the card dimension. No system DMF covers rules like these, so you write a custom DMF with CREATE DATA METRIC FUNCTION. Its first argument is a TABLE with typed columns, and its body is a query that returns the metric.
CREATE OR REPLACE DATA METRIC FUNCTION governance.dmfs.count_positive_numbers( arg_t TABLE( arg_c1 NUMBER, arg_c2 NUMBER, arg_c3 NUMBER ) ) RETURNS NUMBER AS $$ SELECT COUNT(*) FROM arg_t WHERE arg_c1>0 AND arg_c2>0 AND arg_c3>0 $$;Three capabilities take custom DMFs past single-column checks. A custom DMF can accept a second table argument. When you attach it, the table it is set on becomes the first argument and you name the second table by its fully qualified name, which handles referential-integrity, matching and comparison checks across datasets. Inside the body, DATA_METRIC_SCHEDULED_TIME gives the scheduled evaluation time, which is how you write a freshness metric of your own. And expectations work on custom DMFs just as they do on system ones, so the rule's threshold sits next to the association instead of inside the function.
Some thresholds are hard to fix in advance, such as how many rows an hourly load should bring. For those, anomaly detection learns from a DMF's history and flags values above or below a predicted range. You enable it with ANOMALY_DETECTION = TRUE, plus a SENSITIVITY of LOW, MEDIUM or HIGH, when adding the DMF. It currently covers the volume and freshness of your data. It also runs on its own cadence, because the DMF schedule doesn't control how often anomalies are checked.
Checkpoint 8 of 9· Check yourself
Hourly loads into a feature source table vary a lot, so the team turns on ANOMALY_DETECTION = TRUE for the table's row-count DMF instead of fixing a threshold. To catch a drop in volume sooner, an engineer proposes changing DATA_METRIC_SCHEDULE from 60 minutes to 5 minutes. What does that change do to anomaly detection?
DATA_METRIC_SCHEDULE controls how often DMFs and their expectation checks run. It does not control how often Snowflake checks for anomalies, so a shorter schedule won't make anomaly detection respond faster.
“The DMF schedule doesn't affect how often Snowflake checks whether there is an anomaly.”Source: docs.snowflake.com
6.Monitoring feature freshness, schema changes and drift
After ingestion and validation, you need to watch the feature pipelines over time. In Snowsight, AI & ML » Features opens the Feature Store monitoring view for a chosen feature store. It has three tabs, and each tab's queries run on a warehouse you select. The Stream Ingest tab measures the stream ingest API, the endpoint that receives streaming events for stream feature views. That API is served by the feature store's online service, so the tab only has data once the online service is created and running and at least one stream feature view exists.
| Tab | Shows | Shown when |
|---|---|---|
| Offline Refresh | One row per feature view refresh job: Source Data Timestamp, Status, Refresh Duration, Statistics, Dynamic Table link, Query ID | A feature store exists |
| Online Serving | Total Request, Errors (%), Latency p50/p90/p99, scoped to all, a feature group or a feature view | Online serving is enabled for at least one feature view |
| Stream Ingest | Ingest API responses per second, responses by HTTP code, latency p50/p90/p95/p99, concurrency | At least one stream feature view exists |
Freshness lives on the Offline Refresh tab. Each managed feature view is a dynamic table refreshing toward a target lag. The Source Data Timestamp of each refresh tells you how current the feature values actually are, and comparing it with the target lag tells you whether the feature is keeping up. By default the tab hides empty refreshes, those with no changed rows, so select Include empty refreshes if you need to confirm refreshes are happening during quiet periods. A Failed status links to the dynamic table's refresh history for the error. After 5 consecutive scheduled failures, Snowflake suspends the dynamic table, and no refresh runs again until someone resumes it manually. The tab draws on INFORMATION_SCHEMA.DYNAMIC_TABLE_REFRESH_HISTORY, which keeps 7 days. For up to 365 days, with up to 3 hours of latency, query the DYNAMIC_TABLE_REFRESH_HISTORY view directly.
Schema changes show up in more than one place. On the Stream Ingest tab, a rise in 4xx errors points to malformed payloads or a schema mismatch between the payload and the feature view schema. A spike in 429s means ingest is exceeding the provisioned limit. Elevated 5xx errors point to runtime errors or throttling, so check the Concurrency chart for spikes at the same time. A flat or zero ingest rate means the upstream source stopped producing events or the endpoint is unreachable. On the table side, schema evolution in Snowpipe Streaming can add columns silently, and downstream column lineage shows which feature views and models a changed column reaches.
| Symptom | Likely cause |
|---|---|
| High p99 latency | Large payload per request, or a burst of concurrent requests exceeding available capacity |
| Elevated 5xx errors | Runtime errors or resource throttling |
| Elevated 4xx errors | Malformed request payload or a schema mismatch with the feature view schema |
| Flat or zero ingest rate | Upstream source stopped producing events, or the ingest endpoint is unreachable |
Drift is where the native tools stop. The monitoring view reports refresh, serving and ingest health, and anomaly detection covers only volume and freshness, so neither checks whether a feature's distribution has shifted. Feature drift checks therefore go through one of two routes. One is a monitor on the model that consumes the feature (Model monitors (drift, stats and performance metrics)). The other is a custom DMF that computes a distribution statistic and has an expectation on it.
Its status. A Suspended dynamic table runs no new refreshes until someone resumes it manually, and five consecutive scheduled failures cause exactly that. Open the Dynamic Table link to read the failure in the refresh history. Also check that empty refreshes are not simply hidden: select Include empty refreshes.
Checkpoint 9 of 9· Match them up
Match each symptom on the Feature Store Stream Ingest tab to its most likely cause.
Tap a term, then the definition that fits it.
The 4xx family points to the client side: bad payloads or a schema mismatch. A spike in 429s specifically means the provisioned ingest limit is being exceeded, while 5xx errors come from server-side runtime errors or throttling. A flat line means events aren't arriving at all, so the problem is upstream or the endpoint can't be reached.
“Elevated 4xx errors | Malformed request payload or a schema mismatch between the payload and the feature view schema.”Source: docs.snowflake.com
Sources9
Practise it for real
Put a scheduled quality check on an ingestion view and list what depends on it
1.Run ALTER VIEW v1 ADD DATA METRIC FUNCTION SNOWFLAKE.CORE.NULL_COUNT ON (C1) EXPECTATION my_exp ( VALUE < 10);
Why: Pairs a system DMF with an expectation so the measurement turns into a pass/fail check.
You should see: The statement succeeds and the association appears on the object's data quality page in Snowsight.
2.Set the object's DATA_METRIC_SCHEDULE to '5 MINUTE' with ALTER ... SET DATA_METRIC_SCHEDULE = '5 MINUTE';
Why: Without this, the check runs on the default hourly schedule, and only scheduled runs are evaluated.
You should see: After the next five-minute window, a result for my_exp is recorded, with a violation reported if C1 has 10 or more NULLs.
3.Call SNOWFLAKE.CORE.GET_LINEAGE with the object's fully qualified name, 'TABLE' and 'DOWNSTREAM'.
Why: Lists the feature views, datasets and models a schema change to this object would reach.
You should see: One row per downstream edge with DISTANCE values up to 5, or zero rows if no lineage exists. Zero rows is not an error.
Stuck? Get a nudge
If GET_LINEAGE errors instead of returning rows, check that your role has VIEW LINEAGE and a privilege on the object, and that you passed the right domain.
Sources
(9)
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.
“Data movement, such as when data is copied or materialized from one object to another.”
↩︎ Tracing lineage from raw tables to models“Snowsight reveals objects incrementally, one step at a time upstream or downstream from your selection.”
↩︎ Tracing lineage from raw tables to models“You can use the Distance column to determine how far away a column is in the lineage.”
↩︎ Tracing lineage from raw tables to models“If you call a stored procedure anonymously, details about the stored procedure do not appear in the lineage.”
↩︎ Tracing lineage from raw tables to models“A role with the following privileges can access the Lineage tab and view an object’s upstream and downstream lineage objects and dependencies:”
↩︎ Privileges that let a role see lineage“you can grant the REFERENCES privilege on the object.”
↩︎ Privileges that let a role see lineage“By default, the PUBLIC role has this privilege, which means everyone has the ability to view lineage.”
↩︎ Privileges that let a role see lineage“Simply grant the role the RESOLVE ALL privilege on the account”
↩︎ Privileges that let a role see lineage“Objects you don’t have access to are placed in a separate group that is collapsed by default.”
↩︎ Checkpoint - 2.
“The number of levels of lineage to retrieve. The maximum is 5; this is also the default.”
↩︎ Tracing lineage from raw tables to models“The output table contains no rows if no lineage information is available for the specified object; this is not an error.”
↩︎ Tracing lineage from raw tables to models“For ML lineage, use TABLE for feature views (which are dynamic tables and views internally), ‘DATASET’, or ‘MODEL’ for models.”
↩︎ Exam trap 4“For ML lineage, use TABLE for feature views (which are dynamic tables and views internally), ‘DATASET’, or ‘MODEL’ for models.”
↩︎ Prediction - 3.https://docs.snowflake.com/en/user-guide/snowpipe-streaming/data-load-snowpipe-streaming-overviewOfficial docs
“It enables applications to stream rows directly from devices, applications, and services into Snowflake tables or Snowflake-managed Apache Iceberg tables.”
↩︎ Snowpipe Streaming for near real-time features“As low as 5 seconds ingest-to-queryable latency”
↩︎ Snowpipe Streaming for near real-time features“Elastic Channels provide at-least-once delivery without an ordering guarantee.”
↩︎ Snowpipe Streaming for near real-time features“Named Channels provide ordered, exactly-once ingestion within each channel by using offset tokens.”
↩︎ Snowpipe Streaming for near real-time features“Snowflake can add new columns detected in the incoming stream without manual DDL changes.”
↩︎ Snowpipe Streaming for near real-time features“Cleanse, reshape, and transform data during ingestion by using COPY command syntax within the PIPE object.”
↩︎ Snowpipe Streaming for near real-time features“Use Snowpipe when your pipeline already produces files in cloud storage and batch-oriented, higher-latency loading is acceptable.”
↩︎ Exam trap 3“Use Snowpipe Streaming when data arrives as rows from applications, devices, or services and you need low-latency data availability.”
↩︎ Checkpoint - 4.
“An expectation includes a Boolean expression that determines whether the expectation was met or not.”
↩︎ Validating ingestion tables with DMFs and expectations“You can define expectations for both system DMFs and custom DMFs.”
↩︎ Feature validation rules beyond basic checks - 5.
“By default, the DMF schedule runs a DMF once every hour.”
↩︎ Validating ingestion tables with DMFs and expectations“You are not billed for unscheduled data metric function usage, such as calling a DMF with a SELECT statement.”
↩︎ Validating ingestion tables with DMFs and expectations“You cannot set a DMF on a hybrid table or a stream object.”
↩︎ Validating ingestion tables with DMFs and expectations“You can only have 50,000 total associations of DMFs on objects per account.”
↩︎ Validating ingestion tables with DMFs and expectations“Currently, Snowflake can automatically detect anomalies in the volume and freshness of your data.”
↩︎ Feature validation rules beyond basic checks“An expectation is combined with a DMF to create a data quality check.”
↩︎ Key concept“The DMF returns a value based on the current state of your data, but doesn’t define whether that value constitutes a data quality issue”
↩︎ Exam trap 1“The DMF schedule doesn’t affect how often Snowflake checks whether there is an anomaly.”
↩︎ Exam trap 2“The DMF returns a value based on the current state of your data, but doesn’t define whether that value constitutes a data quality issue”
↩︎ Checkpoint“The DMF schedule doesn't affect how often Snowflake checks whether there is an anomaly.”
↩︎ Checkpoint - 6.
“The reclustering of tables doesn’t trigger a DMF to run.”
↩︎ Validating ingestion tables with DMFs and expectations“All data metric functions on a table or view follow the same schedule.”
↩︎ Prediction - 7.
“Resets the schedule for DMFs associated with the object to the default of 60 MINUTE.”
↩︎ Validating ingestion tables with DMFs and expectations“Specifies whether Snowflake uses the DMF to automatically detect anomalies in the table.”
↩︎ Feature validation rules beyond basic checks - 8.
“A user-defined DMF can have more than one argument that accepts a table.”
↩︎ Feature validation rules beyond basic checks“This capability simplifies referential integrity, matching and comparison, or conditional checking across different datasets.”
↩︎ Feature validation rules beyond basic checks“To use the scheduled evaluation time in the body of a custom DMF, for example to define a custom freshness metric, call DATA_METRIC_SCHEDULED_TIME.”
↩︎ Feature validation rules beyond basic checks - 9.
“For Online Serving and Stream Ingest metrics, online service must be created and running.”
↩︎ Monitoring feature freshness, schema changes and drift“The timestamp of the upstream source data that triggered this refresh.”
↩︎ Monitoring feature freshness, schema changes and drift“After 5 consecutive scheduled failures, Snowflake automatically suspends the dynamic table.”
↩︎ Monitoring feature freshness, schema changes and drift“For longer retention (up to 365 days, with up to 3 hours latency), query DYNAMIC_TABLE_REFRESH_HISTORY view directly.”
↩︎ Monitoring feature freshness, schema changes and drift“Malformed request payload or a schema mismatch between the payload and the feature view schema.”
↩︎ Monitoring feature freshness, schema changes and drift“A spike in 429 responses indicates that the ingest rate is exceeding the provisioned limit.”
↩︎ Monitoring feature freshness, schema changes and drift“After 5 consecutive scheduled failures, Snowflake automatically suspends the dynamic table.”
↩︎ Exam trap 5“Elevated 4xx errors | Malformed request payload or a schema mismatch between the payload and the feature view schema.”
↩︎ Checkpoint