Concepts

SQL Processors see data as an independent sequence of infinite events. An Event in this context is a datum of information; the smallest element of information that the underlying system uses to communicate. In Kafka’s case, this is a Kafka record/message.

Two parts of the Kafka record are relevant:

• Key

• Value

These are referred to as facets by the engine. These two components can hold any type of data and Kafka itself is agnostic on the actual storage format for either of these two fields. SQL Processors interpret records as (key, value) pairs, and it exposes ways to manipulate these pairs in several ways.

Queries are applications: SQL Processors

As mentioned above, queries that are meant to be run on streaming data are treated as stand-alone applications. These applications, in the context of the Lenses platform, are referred to as SQL Processors.

A SQL Processor encapsulates a specific Lenses SQL query, its details and everything else Lenses needs to be able to run the query continuously.

Schemas must be available for Structured Data

To support features like:

• Inference of output schemas

• Creation-time validation of input query

• Selections

• Expressions

Lenses SQL Engine Streaming mode needs to have up-to-date schema information for all structured topics that are used as input in a given query. In this context, structured means topics that are using complex storage formats like AVRO or JSON.

INSERT INTO daily-item-purchases-stats
SELECT STREAM
    itemId
    , COUNT(*) AS dailyPurchases
    , AVG(price / quantity) AS average_per_unit
FROM purchases
WINDOW BY TUMBLE 1d
GROUP BY itemId;

For the above query, for example, the purchases topic will need to have a value set to a structured format and a valid schema will need to already have been configured in Lenses. In such schema, fields itemId, price and quantity must be defined, the latter two being of a numerical type.

These requirements ensure the Engine will always be in a position to know what kind of data it will be working with, guaranteeing at the same time that all obvious errors are caught before a query is submitted.

M-N topologies

The UI allows us to visualise any SQL Processor out of the box. For the example:

This visualisation helps to highlight that the Lenses SQL fully supports M-N topologies.

What this means is that multiple input topics can be read at the same time, their data manipulated in different ways and then the corresponding results stored in several output topics, all as part of the same Processor’s topology.

This means that all processing can be done in one go, without having to split parts of a topology into different Processors (which could result in more data being stored and shuffled by Kafka).

Expressions in Lenses SQL

An expression is any part of a Lenses SQL query that can be evaluated to a concrete value (not to be confused with a record value).

In a query like the following:

INSERT INTO target-topic
SELECT STREAM
    CONCAT('a', 'b') AS result1
    , (1 + field1) AS result2
    , field2 AS result3
    , CASE
        WHEN field3 = 'Robert' THEN 'It's bobby'
        WHEN field3 = 'William' THEN 'It's willy'
        ELSE 'Unknown'
      END AS who_is_it
FROM input-topic
WHERE LENGTH(field2) > 5;

CONCAT('a', 'b'), (1 + field1) and field2 are all expressions whose values will be _projected_ onto the output topic, whereas LENGTH(field2) > 5 is an expression whose values will be used to filter out input records.

SQL Processors and Kafka Streams

SQL Processors are built on top of Kafka Streams, and it enriches this tool with an implementation of Lenses SQL that fits well with the architecture and design of Kafka Streams. When executed, they run a Kafka Streams instance.

Consumer groups

Each SQL Processor has an application ID which uniquely identifies it within Lenses. The application ID is used as the Kafka Streams application ID which in turn becomes the underlying Kafka Consumer(s) group identifier.

Scaling

Scaling up or down the number of runners automatically adapts and rebalances the underlying Kafka Streams application in line with the Kafka group semantics.

The advantages of using Kafka Streams as the underlying technology for SQL Processors are several:

• Kafka Streams is an enterprise-ready, widely adopted and understood technology that integrates natively with Kafka

• Using consumer group semantics allows leveraging Kafka’s distribution of workload, fault tolerance and replication out of the box

Data as a flow of events: Streams

A stream is probably the most fundamental abstraction that SQL Processors provide, and it represents an unbounded sequence of independent events over a continuously changing dataset.

Let’s clarify the key terms in the above definition:

• event: an event, as explained earlier, is a datum, that is a (key, value) pair. In Kafka, it is a record.

• continuously changing dataset: the dataset is the totality of all data described by every event received so far. As such, it is changed every time a new event is received.

• unbounded: this means that the number of events changing the dataset is unknown and it could even be infinite

• independent: events don’t relate to each other and, in a stream, they are to be considered in isolation

The main implication of this is that stream transformations (e.g. operations that preserve the stream semantics) are stateless because the only thing they need to take into account is the single event being transformed. Most Projections fall within this category.

Stream Example

To illustrate the meaning of the above definition, imagine that the following two events are received by a stream:

("key1", 10)
("key1", 20)

Now, if the desired operation on this stream was to sum the values of all events with the same key (this is called an Aggregation), the result for "key1" would be 30, because each event is taken in isolation.

Finally, compare this behaviour with that of tables, as explained below, to get an intuition of how these two abstractions are related but different.

Stream syntax

Lenses SQL streaming supports reading a data source (e.g. a Kafka topic) into a stream by using SELECT STREAM.

SELECT STREAM *
FROM input-topic;

The above example will create a stream that will emit an event for each record, including future ones.

Data as a snapshot of the state of the world: Tables

While a stream is useful to have visibility to every change in a dataset, sometimes it is necessary to hold a snapshot of the most current state of the dataset at any given time.

This is a familiar use-case for a database and the Streaming abstraction for this is aptly called table.

For each key, a table holds the latest version received of its value, which means that upon receiving events for keys that already have an associated value, such values will be overridden.

A table is sometimes referred to as a changelog stream, to highlight the fact that each event in the stream is interpreted as an update.

Given its nature, a table is intrinsically a stateful construct, because it needs to keep track of what it has already been seen. The main implication of this is that table transformations will consequently also be stateful, which in this context means that they will require local storage and data being copied.

Additionally, tables support delete semantics. An input event with a given key and a null value will be interpreted as a signal to delete the (key, value) pair from the table.

Finally, a table needs the key for all the input events to not be null. To avoid issues, tables will ignore and discard input events that have a null key.

Table example

To illustrate the above definition, imagine that the following two events are received by a table:

("key1", 10)
("key1", 20)

Now, if the desired operation on this table was to sum the values of all events with the same key (this is called an Aggregation), the result for key1 would be 20, because (key1, 20) is interpreted as an update.

Finally, compare this behaviour with that of streams, as explained above, to get an intuition of how these two abstractions are related but different.

Table syntax

Lenses SQL Streaming supports reading a data source (e.g. a Kafka topic) into a table by using SELECT TABLE.

SELECT TABLE *
FROM input-topic;

The above example will create a table that will treat each event on input-topic, including future ones, as updates.

Tables and compacted topics

Given the semantics of tables, and the mechanics of how Kafka stores data, the Lenses SQL Streaming will set the cleanup.policy setting of every new topic that is created from a table to compact, unless explicitly specified otherwise.

What this means is that the data on the topic will be stored with a semantic more closely aligned to that of a table (in fact, tables in Kafka Streams use compacted topics internally). For further information regarding the implications of this, it is advisable to read the official Kafka Documentation about cleanup.policy.

The duality between streams and tables

Streams and tables have significantly different semantics and use cases, but one interesting observation is that are strongly related nonetheless.

This relationship is known as stream-table duality. It is described by the fact that every stream can be interpreted as a table, and similarly, a table can be interpreted as a stream.

• Stream as Table: A stream can be seen as the changelog of a table. Each event in the stream represents a state change in the table. As such, a table can always be reconstructed by replaying all events of a stream, in order.

• Table as Stream: A table can be seen as a snapshot, at a point in time, of the latest value received for each key in a stream. As such, a stream can always be reconstructed by iterating over each (Key, Value) pair and emitting it as an event.

To clarify the above duality, let’s use a chess game as an example.

On the left side of the above image, a chessboard at a specific point in time during a game is shown. This can be seen as a table where the key is a given piece and the value is its position. Also, on the right-hand side, there is the list of moves that culminated in the positioning described on the left; it should be obvious that this can be seen as a stream of events.

The idea formalised by the stream-table duality is that, as it should be clear from the above picture, we can always build a table from a stream (by applying all moves in order).

It is also always possible to build a stream from a table. In the case of the chess example, a stream could be made where each element represents the current state of a single piece (e.g. w: Q h3).

This duality is very important because it is actively used by Kafka (as well as several other storage technologies), for example, to replicate data and data stores and to guarantee fault tolerance. It is also used to translate table and stream nodes within different parts of a query.

SQL Processors and schemas: a proactive approach

One of the main goals of SQL Processors is to ensure that it uses all the information available to it when a SQL Processor is created to catch problems, suggest improvements and prevent errors. It’s more efficient and less frustrating to have an issue coming up during registration rather than at some unpredictable moment in the future, at runtime, possibly generating corrupted data.

SQL engine will actively check the following during the registration of a processor:

• Validation of all user inputs

• Query lexical correctness

• Query semantics correctness

• Existence of the input topics used within the query

• User permissions to all input and output topics

• Schema alignment between fields and topics used within the query

• Format alignment between data written and output topics, if the latter already exist

When all the above checks pass, the Engine will:

• Generate a SQL Processor able to execute the user’s query

• Generate and save valid schemas for all output topics to be created

• Monitor the processor and make such metrics available to Lenses

The Engine takes a principled and opinionated approach to schemas and typing information; what this means is that, for example, where there is no schema information for a given topic, that topic’s fields will not be available to the Engine, even if they are present in the data; also, if a field in a topic is a string, it will not be possible to use it as a number for example, without explicitly CASTing it.

The Engine’s approach allows it to support naming and reusing parts of a query multiple times. This can be achieved using the dedicated statement WITH.

SET defaults.topic.autocreate=true;
SET commit.interval.ms='1000';
SET enable.auto.commit=false;
SET auto.offset.reset='earliest';

WITH countriesStream AS (
  SELECT STREAM *
  FROM countries
);

WITH merchantsStream AS (
  SELECT STREAM *
  FROM merchants
);


WITH merchantsWithCountryInfoStream AS (
  SELECT STREAM
    m._key AS l_key
    , CONCAT(surname, ', ', name) AS fullname
    , address.country
    , language
    , platform
  FROM merchantsStream AS m
        JOIN countriesStream AS c
            ON m.address.country = c._key  
  WITHIN 1h
);

WITH merchantsCorrectKey AS(
  SELECT STREAM
    l_key AS _key
    , fullname
    , country
    , language
    , platform
  FROM merchantsWithCountryInfoStream
);

INSERT INTO currentMerchants
SELECT STREAM *
FROM merchantsCorrectKey;

INSERT INTO merchantsPerPlatform
SELECT TABLE
  COUNT(*) AS merchants
FROM merchantsCorrectKey
GROUP BY platform;

The WITHs allow for whole sections of the query to be reused and manipulated independently by successive statements, and all this is done by maintaining schema and format alignment and correctness. The reason why this is useful is that it allows to specify queries that split their processing flow without having to redefine parts of the topology. This, in turn, means that less data needs to be read and written to Kafka, improving performance.

This is just an example of what SQL Processors can offer because of the design choices taken and the strict rules implemented at query registration.