Benchmarking Elasticsearch and Rockset: Rockset achieves as much as 4X quicker streaming information ingestion


Rockset is a database used for real-time search and analytics on streaming information. In eventualities involving analytics on large information streams, we’re typically requested the utmost throughput and lowest information latency Rockset can obtain and the way it stacks as much as different databases. To seek out out, we determined to check the streaming ingestion efficiency of Rockset’s subsequent technology cloud structure and evaluate it to open-source search engine Elasticsearch, a preferred sink for Apache Kafka.

For this benchmark, we evaluated Rockset and Elasticsearch ingestion efficiency on throughput and information latency. Throughput measures the speed at which information is processed, impacting the database’s capacity to effectively help high-velocity information streams. Information latency, however, refers back to the period of time it takes to ingest and index the information and make it accessible for querying, affecting the power of a database to offer up-to-date outcomes. We look at latency on the ninety fifth and 99th percentile, provided that each databases are used for manufacturing functions and require predictable efficiency.

We discovered that Rockset beat Elasticsearch on each throughput and end-to-end latency on the 99th percentile. Rockset achieved as much as 4x larger throughput and a couple of.5x decrease latency than Elasticsearch for streaming information ingestion.

On this weblog, we’ll stroll by way of the benchmark framework, configuration and outcomes. We’ll additionally delve underneath the hood of the 2 databases to raised perceive why their efficiency differs relating to search and analytics on high-velocity information streams.

Be taught extra concerning the efficiency of Elasticsearch and Rockset by registering for the tech discuss Evaluating Elasticsearch and Rockset Streaming Ingest and Question Efficiency with CTO Dhruba Borthakur and principal engineer and architect Igor Canadi.

Why measure streaming information ingestion?

Streaming information is on the rise with over 80% of Fortune 100 firms utilizing Apache Kafka. Many industries together with gaming, web and monetary companies are mature of their adoption of occasion streaming platforms and have already graduated from information streams to torrents. This makes it essential to know the size at which finally constant databases Rockset and Elasticsearch can ingest and index information for real-time search and analytics.

In an effort to unlock streaming information for real-time use circumstances together with personalization, anomaly detection and logistics monitoring, organizations pair an occasion streaming platform like Confluent Cloud, Apache Kafka and Amazon Kinesis with a downstream database. There are a number of benefits that come from utilizing a database like Rockset or Elasticsearch together with:

  • Incorporating historic and real-time streaming information for search and analytics
  • Supporting transformations and rollups at time of ingest
  • Very best when information mannequin is in flux
  • Very best when question patterns require particular indexing methods

Moreover, many search and analytics functions are latency delicate, leaving solely a small window of time to take motion. That is the good thing about databases that have been designed with streaming in thoughts, they will effectively course of incoming occasions as they arrive into the system quite than go into sluggish batch processing modes.

Now, let’s bounce into the benchmark so you may have an understanding of the streaming ingest efficiency you may obtain on Rockset and Elasticsearch.

Utilizing RockBench to measure throughput and latency

We evaluated the streaming ingest efficiency of Rockset and Elasticsearch on RockBench, a benchmark that measures the height throughput and end-to-end latency of databases.

RockBench has two parts: an information generator and a metrics evaluator. The information generator writes occasions each second to the database; the metrics evaluator measures the throughput and end-to-end latency or the time from when the occasion is generated till it’s queryable.


Multiple instances of the benchmark connect to the database under test.

A number of cases of the benchmark hook up with the database underneath check.

The information generator generates paperwork, every doc is the dimensions of 1.25KB and represents a single occasion. Which means that 8,000 writes is equal to 10 MB/s.

Peak throughput is the very best throughput at which the database can sustain with out an ever-growing backlog. For this benchmark, we regularly added ingested information in increments of 10 MB/s till the database might not sustainably sustain with the throughput for a interval of 45 minutes. We decided the height throughput because the increment of 10 MB/s above which the database might not maintain the write price.

Every doc has 60 fields containing nested objects and arrays to reflect semi-structured occasions in actual life eventualities. The paperwork additionally include a number of fields which are used to calculate the end-to-end latency:

  • _id: The distinctive identifier of the doc
  • _event_time: Displays the clock time of the generator machine
  • generator_identifier: 64-bit random quantity

The _event_time of that doc is then subtracted from the present time of the machine to reach on the information latency of the doc. This measurement additionally consists of round-trip latency—the time required to run the question and get outcomes from the database again to the shopper. This metric is printed to a Prometheus server and the p50, p95 and p99 latencies are calculated throughout all evaluators.

On this efficiency analysis, the information generator inserts new paperwork to the database and doesn’t replace any present paperwork.

RockBench Configuration & Outcomes

To check the scalability of ingest and indexing efficiency in Rockset and Elasticsearch, we used two configurations with completely different compute and reminiscence allocations. We chosen the Elasticsearch Elastic Cloud cluster configuration that almost all carefully matches the CPU and reminiscence allocations of the Rockset digital cases. Each configurations made use of Intel Ice Lake processors.


Table of the Rockset and Elasticsearch configurations used in the benchmark.

Desk of the Rockset and Elasticsearch configurations used within the benchmark.

The information turbines and information latency evaluators for Rockset and Elasticsearch have been run of their respective clouds and the US West 2 areas for regional compatibility. We chosen Elastic Elasticsearch on Azure as it’s a cloud that gives Intel Ice Lake processors. The information generator used Rockset’s write API and Elasticsearch’s bulk API to write down new paperwork to the databases.

We ran the Elasticsearch benchmark on the Elastic Elasticsearch managed service model v8.7.0, the latest secure model, with 32 major shards, a single duplicate and availability zone. We examined a number of completely different refresh intervals to tune for higher efficiency and landed on a refresh interval of 1 second which additionally occurs to be the default setting in Elasticsearch. We settled on a 32 major shard depend after evaluating efficiency utilizing 64 and 32 shards, following the Elastic steering that shard measurement vary from 10 GB to 50 GB. We ensured that the shards have been equally distributed throughout all the nodes and that rebalancing was disabled.

As Rockset is a SaaS service, all cluster operations together with shards, replicas and indexes are dealt with by Rockset. You possibly can count on to see related efficiency on commonplace version Rockset to what was achieved on the RockBench benchmark.

We ran the benchmark utilizing batch sizes of fifty and 500 paperwork per write request to showcase how properly the databases can deal with larger write charges. We selected batch sizes of fifty and 500 paperwork as they mimic the load sometimes present in incrementally updating streams and excessive quantity information streams.

Throughput: Rockset sees as much as 4x larger throughput than Elasticsearch

Peak throughput is the very best throughput at which the database can sustain with out an ever-growing backlog. The outcomes with a batch measurement of fifty showcase that Rockset achieves as much as 4x larger throughput than Elasticsearch.


Table of the peak throughput and p95 latency of Elasticsearch and Rockset. Databases were evaluated using vCPU 64 and vCPU 128 instances and a batch size of 50.

Desk of the height throughput and p95 latency of Elasticsearch and Rockset. Databases have been evaluated utilizing vCPU 64 and vCPU 128 cases and a batch measurement of fifty.

The outcomes with a batch measurement of fifty showcase that Rockset achieves as much as 4x larger throughput than Elasticsearch.


Table of the peak throughput and p95 latency of Elasticsearch and Rockset. Databases were evaluated using vCPU 64 and vCPU 128 instances and a batch size of 500.

Desk of the height throughput and p95 latency of Elasticsearch and Rockset. Databases have been evaluated utilizing vCPU 64 and vCPU 128 cases and a batch measurement of 500.

With a batch measurement of 500, Rockset achieves as much as 1.6x larger throughput than Elasticsearch.


Graph of the peak throughput of Elasticsearch and Rockset using batches of 50 and 500. Databases were evaluated on 64 and 128 vCPU instances. Higher throughput indicates better performance.

Graph of the height throughput of Elasticsearch and Rockset utilizing batches of fifty and 500. Databases have been evaluated on 64 and 128 vCPU cases. Larger throughput signifies higher efficiency.

One statement from the efficiency benchmark is that Elasticsearch handles bigger batch sizes higher than smaller batch sizes. The Elastic documentation recommends utilizing bulk requests as they obtain higher efficiency than single-document index requests. Compared to Elasticsearch, Rockset sees higher throughput efficiency with smaller batch sizes because it’s designed to course of incrementally updating streams.

We additionally observe that the height throughput scales linearly as the quantity of sources will increase on Rockset and Elasticsearch. Rockset persistently beats the throughput of Elasticsearch on RockBench, making it higher suited to workloads with excessive write charges.

Information Latency: Rockset sees as much as 2.5x decrease information latency than Elasticsearch

We evaluate Rockset and Elasticsearch end-to-end latency on the highest potential throughput that every system achieved. To measure the information latency, we begin with a dataset measurement of 1 TB and measure the common information latency over a interval of 45 minutes on the peak throughput.

We see that for a batch measurement of fifty the utmost throughput in Rockset is 90 MB/s and in Elasticsearch is 50 MB/s. When evaluating on a batch measurement of 500, the utmost throughput in Rockset is 110 MB/s and Elasticsearch is 80 MB/s.


Table of the 50th, 95th and 99th percentile data latencies on batch sizes of 50 and 500 in Rockset and Elasticsearch. Data latencies are recorded for 128 vCPU instances.

Desk of the fiftieth, ninety fifth and 99th percentile information latencies on batch sizes of fifty and 500 in Rockset and Elasticsearch. Information latencies are recorded for 128 vCPU cases.

On the ninety fifth and 99th percentiles, Rockset delivers decrease information latency than Elasticsearch on the peak throughput. What you may also see is that the information latency is inside a tighter certain on Rockset in comparison with the delta between p50 and p99 on Elasticsearch.


Graph of the data latency at 50th, 95th and 99th percentiles at the peak throughput rate of Rockset and Elasticsearch. Shows the results of a batch of 500 on 128 vCPU instances. Lower data latency indicates better performance.

Graph of the information latency at fiftieth, ninety fifth and 99th percentiles on the peak throughput price of Rockset and Elasticsearch. Exhibits the outcomes of a batch of 500 on 128 vCPU cases. Decrease information latency signifies higher efficiency.

Rockset was in a position to obtain as much as 2.5x decrease latency than Elasticsearch for streaming information ingestion.

How did we do it?: Rockset good points attributable to cloud-native effectivity

There have been open questions as as to if it’s potential for a database to attain each isolation and real-time efficiency. The de-facto structure for real-time database techniques, together with Elasticsearch, is a shared nothing structure the place compute and storage sources are tightly coupled for higher efficiency. With these outcomes, we present that it’s potential for a disaggregated cloud structure to help search and analytics on high-velocity streaming information.

One of many tenets of a cloud-native structure is useful resource decoupling, made well-known by compute-storage separation, which gives higher scalability and effectivity. You not have to overprovision sources for peak capability as you may scale up and down on demand. And, you may provision the precise quantity of storage and compute wanted in your utility.

The knock in opposition to decoupled architectures is that they’ve traded off efficiency for isolation. In a shared nothing structure, the tight coupling of sources underpins efficiency; information ingestion and question processing use the identical compute models to make sure that probably the most just lately generated information is accessible for querying. Storage and compute are additionally colocated in the identical nodes for quicker information entry and improved question efficiency.

Whereas tightly coupled architectures made sense prior to now, they’re not needed attributable to advances in cloud architectures. Rockset’s compute-storage and compute-compute separation for real-time search and analytics prepared the ground by isolating streaming ingest compute, question compute and scorching storage from one another. Rockset is ready to guarantee queries entry the latest writes by replicating the in-memory state throughout digital cases, a cluster of compute and reminiscence sources, making the structure well-suited to latency delicate eventualities. Moreover, Rockset creates an elastic scorching storage tier that could be a shared useful resource for a number of functions.


Diagrams of a (a) shared nothing architecture like Elasticsearch and (b) a compute-compute separation architecture introduced by Rockset.

Diagrams of a (a) shared nothing structure like Elasticsearch and (b) a compute-compute separation structure launched by Rockset.

With compute-compute separation, Rockset achieves higher ingest efficiency than Elasticsearch as a result of it solely has to course of incoming information as soon as. In Elasticsearch, which has a primary-backup mannequin for replication, each duplicate must expend compute indexing and compacting newly generated writes. With compute-compute separation, solely a single digital occasion does the indexing and compaction earlier than transferring the newly written information to different cases for utility serving. The effectivity good points from needing to solely course of incoming writes as soon as is why Rockset recorded as much as 4x larger throughput and a couple of.5x decrease end-to-end latency than Elasticsearch on RockBench.

In Abstract: Rockset achieves as much as 4x larger throughput and a couple of.5x decrease latency

On this weblog, we’ve got walked by way of the efficiency analysis of Rockset and Elasticsearch for high-velocity information streams and are available to the next conclusions:

Throughput: Rockset helps larger throughput than Elasticsearch, writing incoming streaming information as much as 4x quicker. We got here to this conclusion by measuring the height throughput, or the speed during which information latency begins monotonically rising, on completely different batch sizes and configurations.

Latency: Rockset persistently delivers decrease information latencies than Elasticsearch on the ninety fifth and 99th percentile, making Rockset properly fitted to latency delicate utility workloads. Rockset gives as much as 2.5x decrease end-to-end latency than Elasticsearch.

Price/Complexity: We in contrast Rockset and Elasticsearch streaming ingest efficiency on {hardware} sources, utilizing related allocations of CPU and reminiscence. We additionally discovered that Rockset gives the most effective worth. For the same value level, you can’t solely get higher efficiency on Rockset however you are able to do away with managing clusters, shards, nodes and indexes. This significantly simplifies operations so your staff can give attention to constructing production-grade functions.

We ran this efficiency benchmark on Rockset’s subsequent technology cloud structure with compute-compute separation. We have been in a position to show that even with the isolation of streaming ingestion compute, question compute and storage Rockset was nonetheless in a position to obtain higher efficiency than Elasticsearch.

If you’re fascinated with studying extra concerning the efficiency of Rockset and Elasticsearch, register for the tech discuss Evaluating Elasticsearch and Rockset Streaming Ingest and Question Efficiency with CTO Dhruba Borthakur and founding engineer and architect Igor Canadi on Might tenth at 9am PDT. They’ll be delving into the efficiency and architectural variations in higher element.

You too can consider Rockset in your personal real-time search and analytics workload by beginning a free trial with $300 in credit. We now have built-in connectors to Confluent Cloud, Kafka and Kinesis together with a bunch of OLTP databases to make it simple so that you can get began.

Authors: Richard Lin, Software program Engineering and Julie Mills, Product Advertising



Related Articles

LEAVE A REPLY

Please enter your comment!
Please enter your name here

Latest Articles