To summarize, in this post I’ve:

Clarified terminology, specifically narrowing the definition of “streaming” to apply to execution engines only, while using more descriptive terms like unbounded data and approximate/speculative results for distinct concepts often categorized under the “streaming” umbrella.

Assessed the relative capabilities of well-designed batch and streaming systems, positing that streaming is in fact a strict superset of batch, and that notions like the Lambda Architecture, which are predicated on streaming being inferior to batch, are destined for retirement as streaming systems mature.

Proposed two high-level concepts necessary for streaming systems to both catch up to and ultimately surpass batch, those being correctness and tools for reasoning about time, respectively.

Established the important differences between event time and processing time, characterized the difficulties those differences impose when analyzing data in the context of when they occurred, and proposed a shift in approach away from notions of completeness and toward simply adapting to changes in data over time.

Looked at the major data processing approaches in common use today for bounded and unbounded data, via both batch and streaming engines, roughly categorizing the unbounded approaches into: time-agnostic, approximation, windowing by processing time, and windowing by event time.

we use a data structure called Resilient Distributed Datasets (RDDs), which keeps data in memory and can recover it without replication by tracking the lineage graph of operations that were used to build it. With RDDs, we show that we can attain sub-second end-to-end latencies. We believe that this is sufficient for many real-world big data applications, where the timescale of the events tracked (e.g., trends in social media) is much higher.

Here are two helpful guidelines (for largely disjoint populations):

If you are going to use a big data system for yourself, see if it is faster than your laptop.
If you are going to build a big data system for others, see that it is faster than my laptop. [...]

We think everyone should have to do this, because it leads to better systems and better research.

Spark 1.2 includes several cross-cutting optimizations focused on performance for large scale workloads. Two new features Databricks developed for our world record petabyte sort with Spark are turned on by default in Spark 1.2. The first is a re-architected network transfer subsystem that exploits Netty 4’s zero-copy IO and off heap buffer management. The second is Spark’s sort based shuffle implementation, which we’ve now made the default after significant testing in Spark 1.1. Together, we’ve seen these features give as much as 5X performance improvement for workloads with very large shuffles.

First and foremost, in Spark 1.1 we introduced a new shuffle implementation called sort-based shuffle (SPARK-2045). The previous Spark shuffle implementation was hash-based that required maintaining P (the number of reduce partitions) concurrent buffers in memory. In sort-based shuffle, at any given point only a single buffer is required. This has led to substantial memory overhead reduction during shuffle and can support workloads with hundreds of thousands of tasks in a single stage (our PB sort used 250,000 tasks).

The optimal histogram is computed using a random-restart hill climbing approximated algorithm.
The algorithm has been shown very fast and accurate: we achieved 99% accuracy compared to an exact dynamic algorithm, with a speed increase of one factor. [...] The amount of information to serve in Voldemort for one year of BBVA's credit card transactions on Spain is 270 GB. The whole processing flow would run in 11 hours on a cluster of 24 "m1.large" instances. The whole infrastructure, including the EC2 instances needed to serve the resulting data would cost approximately $3500/month.