Ingest Transformation

An ingest transformation lets you run a SQL query over the data from your source and only persists the output of that query to the collection. This gives you the power of SQL to drop, rename, or combine fields, filter out incoming rows, and even aggregate incoming documents in real-time with rollups.

Rockset's ingestion platform applies this transformation both during the initial load of a new collection's data, and on an ongoing basis to new documents coming from your source, effectivley giving you a real-time materialized view of your data.

Here we'll cover

• The Structure of an Ingest Transformation Query
• Rollups
• Use Cases For Ingest Transformation
• Example Ingest Transformation Queries

Ingest Transformation Query

You configure an ingest transformation by specifying an ingest transformation query at collection creation time, either in the Rockset Console or through the REST API's create collection endpoint using the field field_mapping_query. To update an ingest transformation after collection creation see below.

The structure of an ingest transformation query is

SELECT

    expression [, ...]

FROM

    _input [ WHERE condition ] [ GROUP BY field [, ...] ] [ HAVING condition ] [ CLUSTER BY field [, ...] ]

Each component above is similar to its normal query counterpart. Here's a brief overview of each clause and how/why you might use it.

SELECT Clause

The SELECT clause is where you define

• Which input fields to include/exclude, e.g. SELECT * or SELECT * EXCEPT (a, b) or SELECT a, b
• Field projections (renaming), e.g. SELECT a AS b
• New expressions to evaluate and add to the final document, e.g. SELECT a + 1 AS b
• Whether object subfields are extracted to top-level fields, e.g. SELECT a, b.*. As with SELECT * you can use the EXCEPT clause to exclude some subfields, e.g. SELECT a, b.* EXCEPT(c)

Note: Explicit projections take precedence over fields from the input document of the same name. For example input document {"a": 1, "b": 2} with ingest transformation SELECT *, a AS b FROM _input will result in document {"a": 1, "b": 1} and the original value of b will be lost.

FROM Clause

Unlike regular queries, ingest transformation queries may only reference the pseudo-collection _input, which is the stream of input documents coming from your source. You may not have any other collection/view/alias in your FROM clause.

However you may use the WITH clause to construct a CTE that can be referenced in the FROM clause.

WHERE Clause

The WHERE clause filters out input documents from your source based on some condition, similar to a normal query.

GROUP BY Clause

The GROUP BY clause is used to construct a rollup collection in which source documents are aggregated at ingestion time and only the resulting summarized aggregates are ingested into your collection. Much like with the GROUP BY clause of a regular SQL query, all input rows with the same values for the fields in the grouping set will be aggregated together to generate one output row.

With rollups you will often want to aggregate along a time dimension, for example saving hourly summaries of sensor readings. In those cases it's best to create a mapping for _event_time and use that as one of the fields in your GROUP BY clause. See the rollups examples below for sample usage.

There are a few key differences however from the GROUP BY clause of a regular query:

1. Fields referenced in the GROUP BY clause of an ingest transformation must be explicit fields from the SELECT clause, such as SELECT a FROM _input GROUP BY a, or in ordinal form, SELECT a FROM _input GROUP BY 1. You cannot do SELECT * FROM _input GROUP BY a, even though you may do this in a regular query.
2. A rollup query is only allowed to contain one grouping set. This means that you can have GROUP BY a, b, c but not GROUP BY GROUPING SETS ((a), (a,b)) as you might do in a regular query.

HAVING Clause

The HAVING clause is used in conjunction with the GROUP BY clause in a rollup query to filter on the aggregated rows generated by the GROUP BY clause and aggregations. Unlike the WHERE clause, which filters out input rows, the HAVING clause only filters out rows after the GROUP BY clause and aggregations have been applied.

As with a normal SQL query, the HAVING clause can refer to the fields in the GROUP BY clause or aggregate functions. For example, if your rollup query groups by fields a and b, then including HAVING a IS NOT NULL in your rollup query will drop the aggregated rows where a is null. If you change the rollup query to contain HAVING COUNT(*) > 1 instead, this keeps the groups where at least two rows were used to form the aggregated row.

CLUSTER BY Clause

The CLUSTER BY clause is used to configure data clustering on a collection. Unlike the other clauses covered above, CLUSTER BY is an ingest transformation only concept and has no analog in regular query execution.

The fields referenced in the CLUSTER BY clause must be explicit fields in the SELECT clause and can be referenced by their name, e.g. SELECT x FROM _input CLUSTER BY x or by their ordinal, e.g. SELECT x FROM _input CLUSTER BY 1.

Rollups

Overview

Rollups are a class of ingest transformation that enable you to aggregate data as it is ingested, combining multiple documents into one.

As new data comes in, Rockset will transform and aggregate it before storing it in your rollup collection. For time-series data, even out-of-order data arrivals that come in after the fact will be properly aggregated automatically.

Using rollups provides two main benefits:

• Reduced storage size because only the aggregated data is stored and indexed.
• Improved query performance because the metrics you care about have been precomputed at ingestion time.

To use rollups, all you need is a GROUP BY clause in your ingest transformation SQL. See the examples below for sample usage.

Rollups are currently available for the following streaming data sources:

• Apache Kafka
• Confluent Cloud
• Confluent Platform (self-managed)
• Amazon Kinesis
• Amazon MSK
• Azure Service Bus
• Azure Event Hubs

Rollups are also available for the following Data Lakes and Warehouses:

• Amazon S3
• Google Cloud Storage
• Azure Blob Storage

Real-time Rollups

Rollups in Rockset are unique and preserve the real-time nature of your data analytics.

Traditionally, rollups are implemented with ahead-of-time aggregation, which means you don’t have access to your data until the end of the aggregation interval. In this approach, aggregation groups are considered immutable, so all data for an aggregation interval must be present for the aggregation to happen.

In contrast, rollups in Rockset allow updates to aggregation groups, so your data remains real-time. You can query your Rockset rollup collection and get up-to-date results as incoming data continues to be aggregated. In addition, because updates to aggregation groups are allowed, out-of-order arrivals "just work" in Rockset.

Ingest Transformation Use Cases

Type Coercion

You can cast fields to enforce type consistency, detect bad records, or extract advanced types from string representations.

Example:

SELECT

    TRY_CAST(last_updated_at AS DATETIME) AS updated_at

FROM

    _input

PII/PHI Masking

If your input dataset has PII (personally identifiable information) or PHI (protected health information) fields, you can use a one-way crypto hash function so that Rockset only stores the hashed value and not the original PII/PHI field.

Example:

SELECT

    TO_HEX(SHA256(email_address)) AS email_hash

FROM

    _input

Precomputing Fields

If your application queries involve complex SQL functions which result in slow query processing, you can pre-compute the result of the expensive SQL expression at ingestion time. Queries on these computed output fields are much faster than executing the SQL expression at query time.

Example:

-- extract the domain from the input email address

SELECT

    REGEXP_EXTRACT_ALL(

        email_address,

        '^([a-zA-Z0-9_.+-]+)@([a-zA-Z0-9-]+\.[a-zA-Z0-9-.]+$)',

        2

    ) AS domain

FROM

    _input

Dropping Fields

If you have unused fields in your input dataset, you can drop them upon data ingestion to save storage space.

Example:

SELECT

    *

EXCEPT

    (large_field_1, large_field_2)

FROM

    _input

Name Standardization

Sometimes the data coming into Rockset can be messy, and you can use the ingest transformation to clean it up.

Example:

SELECT

    mnfct AS manufacturer,

    ser_new_x AS serial_number

FROM

    _input

WHERE

    ser_new_x IS NOT NULL

Special Fields

You can use the ingest transformation to specify the values of the special fields _id, _event_time, and/or _op.

_id

The _id field is the unique identifier of a document in Rockset. If not specified then Rockset will create a new, random id for the document at ingestion time. But if your source data already has a unique identifier, it's best to map it to _id so future updates to the document from the source are properly reflected in Rockset.

Example:

SELECT

    user_id AS _id,

    *

FROM

    _input

Note: If you map a custom value for _id, the value must be non-null and of string type, otherwise the ingestion of the corresponding document will fail.

_event_time

The _event_time field is the document timestamp in Rockset and is used to determine the cutoff for retention.

Example:

SELECT

    TIMESTAMP_MILLIS(created_at_ts) AS _event_time,

    *

FROM

    _input

Note: If you map a custom value for _event_time, the value must be non-null and of either int (microseconds since epoch) or timestamp type, otherwise the ingestion of the corresponding document will fail.

_op

The _op field is used to properly handle CDC operations from your source. The supported operation types (case insensitive) are: INSERT, UPDATE, UPSERT (default), DELETE, REPLACE, and REPSERT. For more on their exact semantics, refer to the documentation for _op.

Example:

SELECT

    CASE

        WHEN _delete = true THEN 'DELETE'

        ELSE 'UPSERT'

    END as _op,

    *

FROM

    _input

Search Tokenization

Text fields in the input can be tokenized at data ingestion time so you can more efficiently run text search queries later.

Example:

SELECT

    TOKENIZE(tweet_text, 'en_US') AS tweet_tokens

FROM

    _input

Importing Geographic Data

You can convert latitude/longitude points or well known text strings to Rockset's native geography type.

Example:

SELECT

    ST_GEOGPOINT(longitude, latitude) AS geo

FROM

    _input

Examples

Let’s look at some more in-depth example ingest transformation queries.

Rollups Examples

Example:

SELECT

    DATE_TRUNC('HOUR', PARSE_TIMESTAMP_ISO8601(ts)) AS _event_time,

    location,

    SUM(cat_count) AS cat_count,

    MAX(cat_count + dog_count) AS peak_animals_seen,

    SQRT(COUNT(*) + AVG(cat_count)) * APPROX_DISTINCT(dog_count) AS my_crazy_metric

FROM

    _input

GROUP BY

    _event_time,

    location

HAVING

    SUM(cat_count) > 5

Imagine the data source is a network of cameras that detect cats and dogs. Every time a cat and/or dog is found, an event is recorded in the form: {"ts": "2021-04-20T16:20:42+0000", "location": "San Francisco", "cat_count": 1, "dog_count": 2}. There is a requirement to count the total number of cats per location, per hour. There is also a requirement to track the locations and times with the highest numbers of animals seen (perhaps to detect if cameras need faster processors that can count more animals). In addition, the my_crazy_metric expression is included here as an example to show just how complex a rollup query expression can get. The HAVING clause keeps the aggregated rows where the total number of cats seen is greater than five.

Example:

SELECT

    DATE_TRUNC(

        'DAY',

        PARSE_TIMESTAMP('%Y/%m/%d %H:%M', datetimeStr)

    ) AS _event_time,

    region,

    SUM(purchasePrice) AS revenue,

    AVG(purchasePrice) AS avgPrice,

    MAX(purchasePrice) AS largestSale,

    SUM(CAST(purchasePrice > 3 AS int)) AS largeSales,

    SUM(purchasePrice * 2) + 1 AS allTheTransforms,

    COUNT(*) AS counts

FROM

    _input

GROUP BY

    _event_time,

    region

In this case, there is raw data for a business’s sales in different regions. Each time a purchase is made, the timestamp, the location of the sale, and the price of the item purchased are recorded. To best understand the business’s financial health, a data-driven approach is taken to aggregate the purchase price data in several different ways. The GROUP BY clause groups data by date and region.

Anonymization and Casting

SELECT

    *

EXCEPT

    -- drop some fields

    (extra_data, more_extra_data, email_address),

    -- type coercion

    TRY_CAST(last_updated_at AS DATETIME) AS last_updated_at,

    -- PII/PHI masking

    TO_HEX(SHA256(email_address)) AS email_address_hash,

FROM

    _input

WHERE

    -- filter input documents

    email_address IS NOT NULL

This query keeps all input fields by default, except for extra_data, more_extra_data, and email_address. It adds two new fields to the final document: last_updated_at and email_address_hash. If there are any input fields named last_updated_at or email_address_hash in the input document, their values in the final document will be the values of the specified expressions, not their values in the input document. It also filters out all input documents where email_address is null.

Data Clustering

SELECT

    cntry AS country,

    cont AS continent,

    capt AS capital,

    gdp_usd as gross_domestic_product,

    ROUND(population / 1e6) AS population_milliions

FROM

    _input

WHERE

    population_millions > 0 CLUSTER BY continent,

    country

This query drops all input fields by default since there is no * in the SELECT clause, and instead it explicitly renames several input fields to their cleaner final names country, continent, capital, gross_domestic_product and computes pouplation_millions from the input field population. The WHERE clause filters out any documents corresponding to countries with a population less than 1M. And data clustering is configured on the collection with clustering key (continent, country) which will speed up queries that filter on continent, or continent and country significantly.

CDC Control Records and _op

SELECT

    user_id AS _id,

    IF(header.delete = true, 'DELETE', 'UPSERT') AS _op,

    IF(

        header.delete = true,

        null,

        PARSE_TIMESTAMP('%Y-%m-%d %H:%M:%E*S', created_at_ts)

    ) AS _event_time,

    *

EXCEPT

    (user_id, header, created_at_ts)

FROM

    _input

This query maps user_id from the source documents to _id in Rockset and interprets _op based on whether a field is set in the header. In the case of a new record coming in, header.delete will be false (or null) and both IF statements will execute their second branch which will set _op = 'UPSERT' and will parse _event_time from _created_at_ts. All fields except the mapped ones and the header will then be inserted into Rockset. When a delete comes through and header.delete is true, then the first branch of the IF statements will execute which will set _op = DELETE and leave _event_time as null. This will issue a delete for the document corresponding to _id.

Precomputing Fields

SELECT

    twitter_handle,

    follower_count,

    -- search tokenization

    TOKENIZE(tweet_text, 'en_US') AS tweet_text_tokens,

    -- functional index

    REGEXP_EXTRACT_ALL(

        email_address,

        '^([a-zA-Z0-9_.+-]+)@([a-zA-Z0-9-]+\.[a-zA-Z0-9-.]+$)',

        2

    ) AS email_address_domain,

    -- native geography type

    ST_GEOGPOINT(longitude, latitude) AS geo,

FROM

    _input

This query drops all input fields by default and only keeps the input fields twitter_handle and follower_count. It also adds three new fields to our final document: tweet_text_tokens, email_address_domain, and geo, each of which can be used to speed up queries that would otherwise have to run expensive operations at query time on the raw input fields.

Using UNDEFINED

SELECT

    *,

    IF(

        user IS NOT NULL,

        CONCAT(user, '@rockset.com'),

        UNDEFINED

    ) AS email_address

FROM

    _input

This query keeps all input fields by default and adds one new field to the final document: email_address.

Note: If user is null, the email_address field will be set to undefined in the final document. Setting a field’s value to undefined in an ingest transformation means that the field will not be present in the final document. Thus, email_address will only be present in the final document if the input document’s user field is not null. See documentation on undefined for more details.

Updating An Ingest Transformation

You can update the ingest transformation for a collection at any time using the REST API's update collection endpoint. Currently this functionality is only available through the API, and ingest transformation updates cannot be done through the Rockset console.

When you update the transformation, you may see a brief pause in ingestion for your sources. It usually only takes a few seconds for the updated transformation to take effect, but with some sources it can take up to a few minutes.

When you update an ingest transformation, only new documents being ingested from your source will have the new ingest transformation applied. The updated ingest transformation will not affect previously ingested documents. There is also no built-in versioning with the ingest transformation to track which documents of a collection had which transformation applied at ingest time. However you can add this in explicitly as part of your transformation, for example

SELECT

    *,

    1 AS version

FROM

    _input

There are some restrictions when updating a transformation:

• You cannot update ingest transformations with rollups configured.
• You cannot add rollups to an existing ingest transformation.
• You cannot modify the clustering of an ingest transformation.

Otherwise you are free to add/remove/change field projections and names, add/remove where clause predicates, etc. To remove the ingest transformation for a collection all-together, update the transformation to the default SELECT * FROM _input.

CSV File Ingestion

When ingesting data from a CSV or TSV file, you can use ingest transformations to fully control how the data is ingested into your collection just like with any other data source. If you don't need to ingest every column from your CSV file, you can specify only the columns you need inside your select clause. To specify the typing of the fields ingested from your file, you should use type casting statements.

As an example, assume you had a CSV file with three fields ('name', 'sign_up_date', and 'age'). An ingest transformation that drops the age field and converts sign_up_date to a datetime type while also renaming the column could be written in the following way:

SELECT

    name,

    TRY_CAST(sign_up_date AS DATETIME) AS signUpDate

FROM

    _input;

For more examples covering further use cases of ingest transformations, review the rest of this document.

Limitations

There are some limitations related to ingest transformation queries that should be kept in mind.

• JOINs are not allowed in the ingest transformation query. When bringing in data from multiple sources into the same collection, the sources are automatically unioned and the result available via the pseudo-collection _input.
• Some SQL clauses including LIMIT, OFFSET, and ORDER BY are unsupported in the ingest transformation.

Source-specific limitations:

• MongoDB– Rockset's MongoDB connector only receives deltas from your MongoDB collection instead of the full document on each update in order to reduce load on your MongoDB cluster. For this reason, with MongoDB any expressions in the ingest transformation query can only involve 1 field. So SELECT a + 1 AS b is allowed, but not SELECT a + b AS c.
• DynamoDB– For collections sourced from DynamoDB, you cannot specify a custom mapping for _id. So SELECT user_id AS _id is invalid. This is so we can properly map updates coming from DynamoDB to the corresponding Rockset document to keep your collection in-sync.