Browse Category: Streaming Storage

Yet Another Cache but for the Streaming World

Traditional cache solutions treat each entry as an immutable blob of data, which poses problems for the append-heavy ingestion workloads that are common in Pravega. Each Event appended to a Stream would either require its own cache entry or need an expensive read-modify-write operation to be included in the Cache. To enable high-performance ingestion of events, big or small, while also providing near-real-time tail reads and high-throughput historical reads, Pravega needs a specialized cache that can natively support the types of workloads that are prevalent in Streaming Storage Systems.

The Streaming Cache, introduced in Pravega with release 0.7, has been designed from the ground up with streaming data in mind and optimizes for appends while organizing the data in a layout that makes eviction and disk spilling easy.

Not all caches are created equal. It is essential to choose a cache that fits the requirements of the system where it will be used, and streaming solutions are no exception to that rule. In this blog post, we describe an innovative way to look at caching that works well with streaming use cases. Continue Reading

Segment Attributes

The ability to pipeline Events to the Segment Store is a key technique that the Pravega Client uses to achieve high throughput, even when dealing with small writes. A Writer appends an Event to its corresponding Segment as soon as it is received, without waiting for previous ones to be acknowledged. To guarantee ordering and exactly once semantics, the Segment Store requires all such appends to be conditional on some known state, which is unique per Writer. This state is stored in each Segment’s Attributes and can be atomically queried and updated with every Segment operation.

Over time, Attributes have evolved to support a variety of use cases, from keeping track of the number of Events in a Segment (enabling auto-scaling) to storing a hash table index. The introduction of Table Segments (key-value stores which contain all of Pravega’s Stream, Transaction and Segment metadata) required the ability to seamlessly manage tens of millions of such Attributes per Segment.

This blog post explains how Segment Attributes work under the hood to provide an efficient key-value store that represents the foundation for several higher-level features. It begins with an overview of how Pravega Writers use them to prevent data duplication or loss and follows up by describing how Segment Attributes are organized as B+Trees in Tier 2 using innovative compaction techniques that reduce write amplification. Continue Reading

Events Big or Small – Bring Them On

Streaming applications typically need to process the events as soon as they arrive. For example, being able to quickly react to events in applications such as fraud detection, manufacturing error detection can result in massive savings. However, due to the limitation of storage systems not being able to handle large numbers of small writes, producers are forced to buffer events before they write. This not only leads to increased latencies from the event generation time to event process time but increases the chance of losing events when writers are not able to store them reliably in failure scenarios.  Pravega has been built from the ground up to be an ideal store for stream processing because not only can it handle frequent small writes well, but it also does that in a consistent and durable way.

This blog explains how Pravega is able to provide excellent throughput using minimal latency for all writes, big or small. It explores the append path of the Segment Store, detailing how we pipeline external requests and use the append-only Tier 1 log to achieve excellent ingestion performance without sacrificing latency. Continue Reading

Streams in and out of Pravega


Reading and writing is the most basic functionality that Pravega offers. Applications ingest data by writing to one or more Pravega streams and consume data by reading data from one or more streams. To implement applications correctly with Pravega, however, it is crucial that the developer is aware of some additional functionality that complements the core write and read calls. For example, writes can be transactional, and readers are organized into groups.

In this post, we cover some basic concepts and functionality that a developer must be aware when developing an application with Pravega, focusing on reads and writes. We encourage the reader to additionally check the Pravega documentation site under the section “Developing Pravega Applications” for some code snippets and more detail. Continue Reading

I have a stream…

Stream On with Pravega

On stage: Pravega

Stream processing is at the spotlight in the space of data analytics, and the reason is rather evident: there is high value in producing insights over continuously generated data shortly rather than wait for it to accumulate and process in a batch. Low latency from ingestion to result is one of the key advantages that stream processing technology offers to its various application domains. In some scenarios, low latency is not only desirable but strictly necessary to guarantee a correct behavior; it makes little sense to wait hours to process a fire alarm or a gas leak event in the Internet of Things.

A typical application to process stream data has a few core components. It has a source that produces events, messages, data samples, which are typically small in size, from hundreds of bytes to a few kilobytes. The source does not have to be a single element, and it can be multiple sensors, mobile terminals, users or servers. . This data is ultimately fed into a stream processor that crunches it and produces intermediate or final output. The final output can have various forms, often being persisted in some data store for consumption (e.g., queries against a data set). Continue Reading

Storage Reimagined for a Streaming World

Driven by the desire to shrink to zero the time it takes to turn massive volumes of raw data into useful information and action, streaming is deceptively simple: just process and act on data as it arrives, quickly, and in a continuous and infinite fashion.

For use cases from Industrial IoT to Connected Cars to Real-Time Fraud Detection and more, we’re increasingly looking to build new applications and customer experiences that react quickly to customer interests and actions, learn and adapt to changing behavior patterns, and the like. But the reality is most of us don’t yet have the tools to do this with production level data volumes, ingestion rates, and fault resiliency. So we do the best we can with bespoke systems piling complexity on top of complexity.

Complexity is symptomatic of fundamental systems design mismatches: we’re using a component for something it wasn’t designed to do, and the mechanisms at our disposal won’t scale from small to large. Continue Reading