Saltar al contenido principal

Streaming Systems

Updated Oct 07, 2020 ·

Real-Time​

Real-time means completing tasks within a guaranteed timeframe based on the process requirements.

  • Response happens within a set period
  • The timeframe can vary: seconds, minutes, hours, or longer
  • Guarantees are part of the system design

Real-Time and Streaming​

Real-time concepts are closely linked to streaming data.

  • Resource-dependent

    • Speed depends on available computing and network resources
    • Limits how fast data can be processed and delivered
    • Data storage affects overall system efficiency

  • Continuous flow

    • Streaming handles data as it arrives
    • Processing speed determines how “real-time” the system feels

Why Scale​

Scaling is about improving system performance to meet different needs.

  • Process the same data faster
  • Handle more data in the same time
  • Deliver data with lower latency
  • Meet guaranteed timeframes (SLAs)

Vertical Scaling​

Vertical scaling improves the performance of a single system by upgrading its components.

  • Faster CPUs or GPUs
  • More or faster RAM
  • Improved networking or storage

Vertical scaling also affects how fast streaming pipelines can process data.

  • Streaming processes continue until data is fully handled
  • SLAs or guarantees require fast processing to meet timeframes
  • Total capacity is limited by system performance

Faster hardware helps pipelines process data quicker, but component choices should be guided by benchmarking and workload analysis.

Horizontal Scaling​

Horizontal scaling means adding more workers to handle more tasks instead of making a single system faster.

  • Scale out by adding more machines
  • Works best with tasks that can be split easily
  • Example: resizing multiple independent images

Horizontal Scaling with Streaming​

Streaming data flows continuously, which can make coordination tricky.

  • Multiple workers require careful communication
  • Coordination adds delays we want to avoid
  • Copying entire pipelines can reduce delays

Streaming works best when each pipeline handles its own data stream without waiting on others.

Pipeline Copies​

Scaling a streaming system often means duplicating pipelines to handle more events.

  • Each pipeline processes events independently
  • Tasks remain contained until complete
  • Vertical scaling within each pipeline is still possible

More pipelines mean faster handling of incoming events and less waiting between tasks.

Additional Considerations​

Other components may be needed to manage a horizontally scaled streaming system.

  • Load balancers direct tasks to pipelines

    • Can use simple round-robin or least-busy strategy
    • Ensures even distribution of events

  • Bottlenecks may appear in parts of the system

    • Disk IO or resource limits can slow processing
    • Consider shortening pipelines
    • Using batch/queue components can help

Communication Issues​

Missing Messages​

Sometimes events never reach the system. This can happen due to network or resource problems.

  • Use sequence numbers to track messages
  • Detect gaps in sequences to request missing messages
  • Re-requesting can slow down the pipeline

Missing messages are tricky but can be managed with proper tracking and retries.

Delayed Messages​

Some events arrive late, affecting processing speed.

  • Messages eventually arrive but later than expected
  • Delays reduce overall system throughput
  • Often caused by upstream resource or network issues

Delayed messages need monitoring to keep the pipeline running smoothly.

Out-of-Order Messages​

Events can arrive in the wrong sequence, mixing older and newer data.

  • Sequence numbers help detect misordered events
  • Some systems reorder events automatically
  • Certain applications (like video/audio) may drop out-of-order events

Handling out-of-order messages ensures correct data processing and integrity.

Repeat Messages​

The same message may appear multiple times due to system issues.

  • Sequence handling can prevent duplicates
  • Some cases allow duplicates without problems
  • Example: repeated sensor readings are often fine

Repeat messages are manageable and sometimes harmless depending on the data type.

Streaming Tools​

Celery​

Celery is a distributed task queue designed to handle jobs in order. It works on single or multiple systems.

  • Handles asynchronous tasks
  • Scales vertically and horizontally

It is useful for real-time task processing and flexible queue management. Sample uses-cases include:

  • Sending emails
  • Resizing images
  • Order processing

Apache Kafka​

Kafka is a distributed event streaming system for sending events between producers and consumers.

  • Producers create events, grouped by topic
  • Consumers handle events differently (logging, transforming, relaying)
  • Stores events based on retention and storage limits

Kafka works well for scalable event transport but requires careful setup. It also supports advanced scenarios in complex systems.

  • Single source of truth for multiple systems
  • Change Data Capture (CDC) for databases
  • Data backups and migrations

Spark Streaming​

Spark Streaming processes data as it arrives, building on Apache Spark.

  • Supports Python, Scala, SQL, and others
  • Handles large-scale data and machine learning tasks
  • Transitions batch processing to stream processing

Spark Streaming focuses on processing and transforming streaming data rather than storing it.