This publish is co-written with Gal Krispel from Riskified.
Riskified is an ecommerce fraud prevention and danger administration platform that helps companies optimize on-line transactions by distinguishing reliable clients from fraudulent ones.
Utilizing synthetic intelligence and machine studying (AI/ML), Riskified analyzes real-time transaction knowledge to detect and forestall fraud whereas maximizing transaction approval charges. The platform supplies a chargeback assure, defending retailers from losses attributable to fraudulent transactions. Riskified’s options embrace account safety, coverage abuse prevention, and chargeback administration software program, making it a complete device for decreasing danger and enhancing buyer expertise. Companies throughout numerous industries, together with retail, journey, and digital items, use Riskified to extend income whereas minimizing fraud-related losses. Riskified’s core enterprise of real-time fraud prevention makes low-latency streaming applied sciences a basic a part of its resolution.
Companies typically can’t afford to attend for batch processing to make vital selections. With real-time knowledge streaming applied sciences like Apache Flink, Apache Spark, and Apache Kafka Streams, organizations can react immediately to rising tendencies, detect anomalies, and improve buyer experiences. These applied sciences are highly effective processing engines that carry out analytical operations at scale. Nevertheless, unlocking the complete potential of streaming knowledge typically requires complicated engineering efforts, limiting accessibility for analysts and enterprise customers.
Streaming pipelines are in excessive demand from Riskified’s Engineering division. Subsequently, a user-friendly interface for creating streaming pipelines is a vital function to extend analytical precision for detecting fraudulent transactions.
On this publish, we current Riskified’s journey towards enabling self-service streaming SQL pipelines. We stroll by the motivations behind the shift from Confluent ksqlDB to Apache Flink, the structure Riskified constructed utilizing Amazon Managed Service for Apache Flink, the technical challenges they confronted, and the options that helped them make streaming accessible, scalable, and production-ready.
Utilizing SQL to create streaming pipelines
Clients have a spread of open supply knowledge processing applied sciences to select from, similar to Flink, Spark, ksqlDB, and RisingWave. Every platform presents a streaming API for knowledge processing. SQL streaming jobs provide a strong and intuitive method to course of real-time knowledge with minimal complexity. These pipelines use SQL, a extensively recognized and declarative language, to carry out real-time transformations, filtering, aggregations, and joins in steady knowledge streams.
As an instance the ability of streaming SQL in ecommerce fraud prevention, contemplate the idea of velocity checks, that are a vital fraud detection sample. Velocity checks are a sort of safety measure used to detect uncommon or speedy exercise by monitoring the frequency and quantity of particular actions inside a given timeframe. These checks assist determine potential fraud or abuse by analyzing repeated behaviors that deviate from regular consumer patterns. Widespread examples embrace detecting a number of transactions from the identical IP deal with in a short while span, monitoring bursts of account creation makes an attempt, or monitoring the repeated use of a single fee technique throughout completely different accounts.
Use case: Riskified’s velocity checks
Riskified carried out a real-time velocity examine utilizing streaming SQL to watch buying habits based mostly on consumer identifier.
On this setup, transaction knowledge is repeatedly streamed by a Kafka subject. Every message accommodates consumer agent data originating from the browser, together with the uncooked transaction knowledge. Streaming SQL queries are used to mixture the variety of transactions originating from a single consumer identifier inside brief time home windows.
For instance, if the variety of transactions from a given consumer identifier exceeds a sure threshold inside a 10-second interval, this would possibly sign fraudulent exercise. When that threshold is breached, the system can routinely flag or block the transactions earlier than they’re accomplished. The next determine and accompanying code present a simplified instance of the streaming SQL question used to detect this habits.
SELECT userIdentifier,TUMBLE_START(createdAt, INTERVAL ’10’ SECONDS)
AS windowStart,TUMBLE_END(createdAt, INTERVAL ’10’ SECONDS)
AS windowEnd, COUNT(*) AS paymentAttempts
FROM transactions
WINDOW TUMBLING (SIZE 10 SECONDS)
GROUP BY userIdentifier;
Though defining SQL queries over static datasets would possibly seem easy, creating and sustaining sturdy streaming functions introduces distinctive challenges. Conventional SQL operates on bounded datasets, that are finite collections of information saved in tables. In distinction, streaming SQL is designed to course of steady, unbounded knowledge streams resembling the SQL syntax.
To deal with these challenges at scale and make streaming job creation accessible to engineering groups, Riskified carried out a self-serve resolution based mostly on Confluent ksqlDB, utilizing its SQL interface and built-in Kafka integration. Engineers might outline and deploy streaming pipelines utilizing SQL, chaining ksqlDB streams from supply to sink. The system supported each stateless and stateful processing instantly on Kafka subjects, with Avro schemas used to outline the construction of streaming knowledge.
Though ksqlDB supplied a quick and approachable place to begin, it will definitely revealed a number of limitations. These included challenges with schema evolution, difficulties in managing compute sources, and the absence of an abstraction for managing pipelines as a cohesive unit. In consequence, Riskified started exploring different applied sciences that might higher help its increasing streaming use circumstances. The next sections define these challenges in additional element.
Evolving the stream processing structure
In evaluating alternate options, Riskified centered on applied sciences that might deal with the precise calls for of fraud detection whereas preserving the simplicity that made the unique method interesting. The workforce encountered the next challenges in sustaining the earlier resolution:
Schemas are managed in Confluent Schema Registry, and the message format is Avro with FULL compatibility mode enforced. Schemas are continuously evolving in line with enterprise necessities. They’re model managed utilizing Git with a strict steady integration and steady supply (CI/CD) pipeline. As schemas grew extra complicated, ksqlDB’s method to schema evolution didn’t routinely incorporate newly added fields. This habits required dropping streams and recreating them so as to add new fields as an alternative of simply restarting the applying to include new fields. This method prompted inconsistencies with offset administration as a result of stream’s tear-down.
ksqlDB enforces a TopicNameStrategy schema registration technique, which supplies 1:1 schema-to-topic coupling. This implies the precise schema definition needs to be registered a number of instances, one time for every subject it’s used for. Riskified’s schema registry deployment makes use of RecordNameStrategy for schema registration. It’s an environment friendly schema registry technique that enables for sharing schemas throughout a number of subjects, storing fewer schemas, and decreasing registry administration overhead. Having blended methods within the schema registry prompted errors with Kafka shopper shoppers trying to decode messages, as a result of the shopper implementation anticipated a RecordNameStrategy in line with Riskified’s commonplace.
ksqlDB internally registers schema definitions in particular methods the place fields are interpreted as nullable, and Avro Enum varieties are transformed to Strings. This habits prompted deserialization errors when trying emigrate native Kafka shopper functions to make use of the ksqlDB output subject. Riskified’s code base makes use of the Scala programming language, the place non-compulsory fields within the schema are interpreted as Choice. Reworking each discipline as non-compulsory within the schema definition required heavy refactoring, treating all Enum fields as Strings, and dealing with the Choice knowledge kind for each discipline that requires secure dealing with. This cascading impact made the migration course of extra concerned, requiring further time and sources to realize a clean transition.
Managing useful resource rivalry in ksqlDB streaming workloads
ksqlDB queries are compiled right into a Kafka Streams topology. The question definition defines the topology’s habits.
Streaming question sources are shared fairly than remoted. This method sometimes results in the overallocation of cluster sources. Its duties are distributed throughout nodes in a ksqlDB cluster. This structure means processing duties with no useful resource isolation, and a selected process can affect different duties operating on the identical node.
Useful resource rivalry between duties on the identical node is frequent in a production-intensive atmosphere when utilizing a cluster structure resolution. Operation groups typically fine-tune cluster configurations to keep up acceptable efficiency, regularly mitigating points by over-provisioning cluster nodes.
Challenges with ksqlDB pipelines
A ksqlDB pipeline is a sequence of particular person streams and lacks flow-level abstraction. Think about a fancy pipeline the place a shopper publishes to a number of subjects. In ksqlDB, every subject (each enter and output) have to be managed as a separate stream abstraction. Nevertheless, there is no such thing as a high-level abstraction to characterize a whole pipeline that chains these streams collectively. In consequence, engineering groups should manually assemble particular person streams right into a cohesive knowledge stream, with out built-in help for managing them as a single, full pipeline.
This architectural method notably impacts operational duties. Troubleshooting requires analyzing every stream individually, making it tough to watch and keep pipelines that include dozens of interconnected streams. When points happen, the well being of every stream must be checked individually, with no logical knowledge stream part to assist perceive the relationships between streams or their position within the total pipeline. The absence of a unified view of the information stream considerably elevated operational complexity.
Flink in its place
Riskified started exploring alternate options for its streaming platform. The necessities have been clear: a powerful processing know-how that mixes a wealthy low-level API and a streaming SQL engine, backed by a powerful open supply group, confirmed to carry out in essentially the most demanding manufacturing environments.
In contrast to the earlier resolution, which supported solely Kafka-to-Kafka integration, Flink presents an array of connectors for numerous databases and Streaming platforms. It was rapidly acknowledged that Flink had the potential to deal with complicated streaming use circumstances.
Flink presents a number of deployment choices, together with standalone clusters, native Kubernetes deployments utilizing operators, and Hadoop YARN clusters. For enterprises in search of a completely managed choice, cloud suppliers like AWS provide managed Flink companies that assist alleviate operational overhead, similar to Managed Service for Apache Flink.
Advantages of utilizing Managed Service for Apache Flink
Riskified determined to implement an answer utilizing Managed Service for Apache Flink. This selection provided a number of key benefits:
It presents a fast and dependable method to run Flink functions and reduces the operational overhead of independently managing the infrastructure.
Managed Service for Apache Flink supplies true job isolation by operating every streaming software in its devoted cluster. This implies you possibly can handle sources individually for every job and scale back the chance of heavy streaming jobs inflicting useful resource hunger for different operating jobs.
It presents built-in monitoring utilizing Amazon CloudWatch metrics, software state backup with managed snapshots, and computerized scaling.
AWS presents complete documentation and sensible examples to assist speed up the implementation course of.
With these options, Riskified might give attention to what really issues—getting nearer to the enterprise objective and beginning to write functions.
Utilizing Flink’s streaming SQL engine
Builders can use Flink to construct complicated and scalable streaming functions, however Riskified noticed it as greater than only a device for consultants. They wished to democratize the ability of Flink right into a device for all the group, to unravel complicated enterprise challenges involving real-time analytics necessities with no need a devoted knowledge skilled.
To exchange their earlier resolution, they envisioned sustaining a “construct as soon as, deploy many” software, which encapsulates the complexity of the Flink programming and permits the customers to give attention to the SQL processing logic.
Kafka was maintained because the enter and output know-how for the preliminary migration use case, which has similarities to the ksqlDB setup. They designed a single, versatile Flink software the place end-users can modify the enter subjects, SQL processing logic, and output locations by runtime properties. Though ksqlDB primarily focuses on Kafka integration, Flink’s in depth connector ecosystem permits it to increase to numerous knowledge sources and locations in future phases.
Managed Service for Apache Flink supplies a versatile method to configure streaming functions with out modifying their code. Through the use of runtime parameters, you possibly can change the applying’s habits with out modifying its supply code.
Utilizing Managed Service for Apache Flink for this method consists of the next steps:
Apply parameters for the enter/output Kafka subject, a SQL question, and the enter/output schema ID (assuming you’re utilizing Confluent Schema Registry).
Use AvroSchemaConverter to transform an Avro schema right into a Flink desk.
Apply the SQL processing logic and save the output as a view.
Sink the view outcomes into Kafka.
The next diagram illustrates this workflow.
Performing Flink SQL question compilation and not using a Flink runtime atmosphere
Offering end-users with vital management to outline their pipelines makes it vital to confirm the SQL question outlined by the consumer earlier than deployment. This validation prevents failed or hanging jobs that might devour pointless sources and incur pointless prices.
A key problem was validating Flink SQL queries with out deploying the complete Flink runtime. After investigating Flink’s SQL implementation, Riskified found its dependency on Apache Calcite – a dynamic knowledge administration framework that handles SQL parsing, optimization, and question planning independently of information storage. This perception enabled utilizing Calcite instantly for question validation earlier than job deployment.
You should know the way the information is structured to validate a Flink SQL question on a streaming supply like a Kafka subject. In any other case, surprising errors would possibly happen when trying to question the streaming supply. Though an anticipated schema is used with relational databases, it’s not enforced for streaming sources.
Schemas assure a deterministic construction for the information saved in a Kafka subject when utilizing a schema registry. A schema could be materialized right into a Calcite desk that defines how knowledge is structured within the Kafka subject. It permits inferring desk buildings instantly from schemas (on this case, Avro format was used), enabling thorough field-level validation, together with kind checking and discipline existence, all earlier than job deployment. This desk can later be used to validate the SQL question.
The next code is an instance of supporting fundamental discipline varieties validation utilizing Calcite’s AbstractTable:
public class FlinkValidator {
public static void validateSQL(String sqlQuery, Schema avroSchema) throws Exception {
SqlParser.Config sqlConfig = SqlParser.config()
.withCaseSensitive(true);
SqlParser sqlParser = SqlParser.create(sqlQuery, sqlConfig);
SqlNode parsedQuery = sqlParser.parseQuery();
RelDataTypeFactory typeFactory = new SqlTypeFactoryImpl(RelDataTypeFactory.DEFAULT);
CalciteSchema rootSchema = createSchemaWithAvro(avroSchema);
SqlValidator validator = SqlValidatorUtil.newValidator(
Frameworks.newConfigBuilder().construct().getOperatorTable(),
rootSchema.createCatalogReader(Collections.emptyList(), typeFactory),
typeFactory,
SqlValidator.Config.DEFAULT
);
validator.validate(parsedQuery);
}
personal static CalciteSchema createSchemaWithAvro(Schema avroSchema) {
CalciteSchema rootSchema = CalciteSchema.createRootSchema(true);
rootSchema.add(“TABLE”, new SimpleAvroTable(avroSchema));
return rootSchema;
}
personal static class SimpleAvroTable extends org.apache.calcite.schema.impl.AbstractTable {
personal last Schema avroSchema;
public SimpleAvroTable(Schema avroSchema) {
this.avroSchema = avroSchema;
}
@Override
public RelDataType getRowType(RelDataTypeFactory typeFactory) {
RelDataTypeFactory.Builder builder = typeFactory.builder();
for (Schema.Subject discipline : avroSchema.getFields()) {
builder.add(discipline.title(), convertAvroType(discipline.schema(), typeFactory));
}
return builder.construct();
}
personal RelDataType convertAvroType(Schema schema, RelDataTypeFactory typeFactory) {
swap (schema.getType()) {
case STRING:
return typeFactory.createSqlType(SqlTypeName.VARCHAR);
case INT:
return typeFactory.createSqlType(SqlTypeName.INTEGER);
default:
return typeFactory.createSqlType(SqlTypeName.ANY);
}
}
}
}
You may combine this validation method as an intermediate step earlier than creating the applying. You may create a streaming job programmatically with the AWS SDK, AWS Command Line Interface (AWS CLI), or Terraform. The validation happens earlier than submitting the streaming job.
Flink SQL and Confluent Avro knowledge kind mapping limitation
Flink supplies a number of APIs designed for various ranges of abstraction and consumer experience:
Flink SQL sits on the highest stage, permitting customers to precise knowledge transformations utilizing acquainted SQL syntax, which is right for analysts and groups snug with relational ideas.
The Desk API presents an identical method however is embedded in Java or Python, enabling type-safe and extra programmatic expressions.
For extra management, the DataStream API exposes low-level constructs to handle occasion time, stateful operations, and complicated occasion processing.
On the most granular stage, the ProcessFunction API supplies full entry to Flink’s runtime options. It’s appropriate for superior use circumstances that demand detailed management over state and processing habits.
Riskified initially used the Desk API to outline streaming transformations. Nevertheless, when deploying their first Flink job to a staging atmosphere, they encountered serialization errors associated to the avro-confluent library and Desk API. Riskified’s schemas rely closely on Avro Enum varieties, which the avro-confluent integration doesn’t totally help. In consequence, Enum fields have been transformed to Strings, resulting in mismatches throughout serialization and errors when trying to sink processed knowledge again to Kafka utilizing Flink’s Desk API.
Riskified developed another method to beat the Enum serialization limitations whereas sustaining schema necessities. They found that Flink’s DataStream API might accurately deal with Confluent’s Avro information serialization with Enum fields, in contrast to the Desk API. They carried out a hybrid resolution combining each APIs as a result of the pipeline solely required SQL processing on the supply Kafka subject. It could sink to the output with none further processing. The Desk API is used for knowledge processing and transformations, solely changing to the DataStream API on the last output stage.
Managed Service for Apache Flink helps Flink APIs. It could swap between the Desk API and the DataStream API.
A MapFunction can convert the Row kind of the Desk API right into a DataStream of GenericRecord. The MapFunction maps Flink’s Row knowledge kind into GenericRecord varieties by iterating over the Avro schema fields and constructing the GenericRecord from the Flink Row kind, casting the Row fields into the proper knowledge kind in line with the Avro schema. This conversion is required to beat the avro-confluent library limitation with Flink SQL.
The next diagram and illustrates this workflow.
The next code is an instance question:
// SQL Question for filtering
Desk queryResults = tableEnv.sqlQuery(
“SELECT * FROM InputTable”);
// 1. Convert question outcomes from Desk API to a DataStream and use DataStream API to sink question outcomes to Kafka subject
DataStream rowStream = tableEnv.toDataStream(queryResults);
// Fetch the schema string from the schema registry
String schemaString = fetchSchemaString(schemaRegistryURL, schemaSubjectName);
// 2. Convert Row to GenericRecord with express TypeInformation, utilizing customized AvroMapper
TypeInformation typeInfo = new GenericRecordAvroTypeInfo(avroSchema);
DataStream genericRecordStream = rowStream
.map(new AvroMapper(schemaString))
.returns(typeInfo); // Explicitly set TypeInformation
// 3. Outline Kafka sink utilizing ConfluentRegistryAvroSerializationSchema
KafkaSink kafkaSink = KafkaSink.builder()
.setBootstrapServers(bootstrapServers)
.setRecordSerializer(
KafkaRecordSerializationSchema.builder()
.setTopic(sinkTopic)
.setValueSerializationSchema(
ConfluentRegistryAvroSerializationSchema.forGeneric(
schemaSubjectName,
avroSchema,
schemaRegistryURL
)
)
.construct()
)
.construct();
// Sink to Kafka
genericRecordStream.sinkTo(kafkaSink);
CI/CD With Managed Service for Apache Flink
With Managed Service for Apache Flink, you possibly can run a job by deciding on an Amazon Easy Storage Service (Amazon S3) key containing the applying JAR. Riskified’s Flink code base was structured as a multi-module repository to help further use circumstances moreover supporting self-service SQL. Every Flink job supply code within the repository is an impartial Java module. The CI pipeline carried out a strong construct and deployment course of consisting of the next steps:
Construct and compile every module.
Run checks.
Package deal the modules.
Add the artifact to the artifacts bucket twice: one JAR below -.jar and the second as -latest.jar, resembling a Docker registry like Amazon Elastic Container Registry (Amazon ECR). Managed Service for Apache Flink jobs makes use of the most recent tag artifact on this case. Nevertheless, a replica of previous artifacts is saved for code rollback causes.
A CD course of follows this course of:
When merged, it lists all jobs for every module utilizing the AWS CLI for Managed Service for Apache Flink.
The appliance JAR location is up to date for every software, which triggers a deployment.
When the applying is in a operating state with no errors, the next software shall be continued.
To permit secure deployment, this course of is completed progressively for each atmosphere, beginning with the staging atmosphere.
Self-service interface for submitting SQL jobs
Riskified believes an intuitive UI is essential for system adoption and effectivity. Nevertheless, creating a devoted UI for Flink job submission requires a realistic method, as a result of it won’t be price investing in except there’s already an internet interface for inner growth operations.
Investing in UI growth ought to align with the group’s present instruments and workflows. Riskified had an inner internet portal for comparable operations, which made the addition of Flink job submission capabilities a pure extension of the self-service infrastructure.
An AWS SDK was put in on the net server to permit interplay with AWS elements. The shopper receives consumer enter from the UI and interprets it into runtime properties to regulate the habits of the Flink software. The online server then makes use of the CreateApplication API motion to submit the job to Managed Service for Apache Flink.
Though an intuitive UI considerably enhances system adoption, it’s not the one path to accessibility. Alternatively, a well-designed CLI device or REST API endpoint can present the identical self-service capabilities.
The next diagram illustrates this workflow.
Manufacturing expertise: Flink’s implementation upsides
The transition to Flink and Managed Service for Apache Flink proved environment friendly in quite a few points:
Schema evolution and knowledge dealing with – Riskified can both periodically fetch up to date schemas or restart functions when schemas evolve. They’ll use present schemas with out self-registration.
Useful resource isolation and administration – Managed Service for Apache Flink runs every Flink job as an remoted cluster, decreasing useful resource rivalry between jobs.
Useful resource allocation and cost-efficiency – Managed Service for Apache Flink permits minimal useful resource allocation with computerized scaling, proving to be extra cost-efficient.
Job administration and stream visibility – Flink supplies a cohesive knowledge stream abstraction by its job and process mannequin. It manages all the knowledge stream in a single job and distributes the workload evenly over a number of nodes. This unified method permits higher visibility into all the knowledge pipeline, simplifying monitoring, troubleshooting, and optimizing complicated streaming workflows.
Constructed-in restoration mechanism – Managed Service for Apache Flink routinely creates checkpoints and savepoints that allow stateful Flink functions to recuperate from failures and resume processing with out knowledge loss. With this function, streaming jobs are sturdy and might recuperate safely from errors.
Complete observability – Managed Service for Apache Flink exposes CloudWatch metrics that monitor Flink software efficiency and statistics. It’s also possible to create alarms based mostly on these metrics. Riskfied determined to make use of the Cloudwatch Prometheus Exporter to export these metrics to Prometheus and construct PrometheusRules to align Flink’s monitoring to the Riskified commonplace, which makes use of Prometheus and Grafana for monitoring and alerting.
Subsequent steps
Though the preliminary focus was Kafka-to-Kafka streaming queries, Flink’s wide selection of sink connectors presents the potential for pluggable multi-destination pipelines. This versatility is on Riskfied’s roadmap for future enhancements.
Flink’s DataStream API supplies capabilities that reach far past self-serving streaming SQL capabilities, opening new avenues for extra subtle fraud detection use circumstances. Riskified is exploring methods to make use of DataStream APIs to boost ecommerce fraud prevention methods.
Conclusions
On this publish, we shared how Riskified efficiently transitioned from ksqlDB to Managed Service for Apache Flink for its self-serve streaming SQL engine. This transfer addressed key challenges like schema evolution, useful resource isolation, and pipeline administration. Managed Service for Apache Flink presents options similar to together with remoted jobs environments, computerized scaling, and built-in monitoring, which proved extra environment friendly and cost-effective. Though Flink SQL limitations with Kafka required workarounds, utilizing Flink’s DataStream API and user-defined capabilities resolved these points. The transition has paved the way in which for future growth with multi-targets and superior fraud detection capabilities, solidifying Flink as a strong and scalable resolution for Riskified’s streaming wants.
If Riskified’s journey has sparked your curiosity in constructing a self-service streaming SQL platform, right here’s methods to get began:
Be taught extra about Managed Service for Apache Flink:
Get hands-on expertise:
In regards to the authors
Gal Krispel is a Information Platform Engineer at Riskified, specializing in streaming applied sciences similar to Apache Kafka and Apache Flink. He focuses on constructing scalable, real-time knowledge pipelines that energy Riskified’s core merchandise. Gal is especially inquisitive about making complicated knowledge architectures accessible and environment friendly throughout the group. His work spans real-time analytics, event-driven design, and the seamless integration of stream processing into large-scale manufacturing techniques.
Sofia Zilberman works as a Senior Streaming Options Architect at AWS, serving to clients design and optimize real-time knowledge pipelines utilizing open-source applied sciences like Apache Flink, Kafka, and Apache Iceberg. With expertise in each streaming and batch knowledge processing, she focuses on making knowledge workflows environment friendly, observable, and high-performing.
Lorenzo Nicora works as Senior Streaming Answer Architect at AWS, serving to clients throughout EMEA. He has been constructing cloud-centered, data-intensive techniques for over 25 years, working throughout industries each by consultancies and product corporations. He has used open-source applied sciences extensively and contributed to a number of tasks, together with Apache Flink, and is the maintainer of the Flink Prometheus connector.
