Camunda BPM 7.2: Performance and Scalability (English)

Hands-on Webinar

Camunda BPM 7.2

Performance and Scalability

Daniel Meyer

Process Engine Expert

Technical Project Lead

@meyerdan | [email protected]

Bernd Rücker

10+ years experience with workflow and Java

Co-Founder of Camunda

Evangelist & Head of Consulting

@berndruecker | [email protected]

Your speakers today

Performance is a difficult topic

It always depends

−On hardware

−On software environment (OS, Java, App Server, Database, …)

−On Service Tasks in the process

−On network topology (e.g. remote database, web services, …)

−On concurrent requests, database load, …

There is no simple answer to performance

But we always succeed – in each and every real-life situation

−Handling millions of process instances / day

−Handling more than 1.000 process instances / second

−Handling thousands of parallel users

Performance is a difficult topic

We are much faster than competition

see http://camunda.com/landing/whitepaper-camunda-jbpm/

In our tests, Camunda‘s throughput was 10x – 30x higher than with JBoss jBPM.

1. Understand basic engine architecture

2. Understand influence parameters on performance

3. Discuss performance improvement approaches

4. See example figures / measurements

5. Discuss future scenarios (e.g. sharding, NoSQL, …)

What we do today

Basic Engine Architecture

We use Optimistic Locking

Runtime vs. History Database

Runtime database schema

Learning #1: The architecture it damn simple – and the bottleneck is not the process engine!

Biggest influence on Performance

Database Delegation Code

Call Service

Clustering via shared database

Learning #2: All state is in the database so clustering gets really easy. camunda scales! More on this later…

„But what can I do if performance IS a problem?“

1. Tasklist

2. (History) Queries

3. Job Execution

Typical Areas of performance issues

Process/Task Variables

−Show in list

−Use in Search/Filter

Support for Pagination

Big number of users accessing the tasklist very often

Implementation challenge

Provide a generic database schema

Complex data types are serialized – no SQL-JOIN possible

Variables are stored in one row per variable – multiple SQL-JOINs might be required

Some customers use 10-30 variables

Tasklist Requirements

Add Process Variables optimized (and only used) for Queries

−Extract attributes

−Combine variables to work with LIKE

Use own queries

−Native – if you want to improve the WHERE

−Custom – if you want to SELECT multiple information at once

Own TaskInfo or ProcessInstanceInfo entities

−Persisted as MyBatis or JPA entities

−Combine all attributes – allow to query tasks without (or with one) JOIN only

−Synchronisation via Listener – or use ProcessInstanceInfo as single source

Solution Approaches: Tasklist

Example

Customer

- customerId - company - …

Your DB camunda

PROCESS_VARIABLES

customerId ... searchField

4711 ... 4711#camunda#Berlin#...

1

2

Native Query:

3

Custom Query:

4

Java API – results are

camunda „Task“ entities

Own MyBatis mapping – result can be anything.

Called via custom code.

Example

TaskInfo

- taskId - customerId - companyName - contractId - productName - …

Your DB

camunda

PROCESS_VARIABLES

customerId contractId productId

4711 0815 42

5

TaskInfo Entity (or ProcessInstanceInfo)

The challenge:

−Indexes cost space and performance in writing data

−We provide a generic database schema without knowing what you exactly do with it

−We constantly work on the right balance of too many and too less indexes

What you can do:

−Check indexes and slow query log

−Add index where appropriate for your situation (perfectly OK with us, you do not loose support!)

−As Enterprise Customer you can always discuss/validate changes with support

Example: create index PROC_DEF_ID_END_TIME ON ACT_HI_PROCINST (PROC_DEF_ID_,END_TIME_)

(History) Queries

You can also customize history

Custom History (e.g.

ElasticSearch)

Different History Levels: - NONE - ACTIVITY - AUDIT - FULL - CUSTOM (own Filter written

in Java, e.g. „only variable X“, „not process Y“, …)

Example for custom log level: https://github.com/camunda/camunda-bpm-examples/tree/master/process-engine-plugin/custom-history-level

Job Execution

Asynchronous Continuation involve Jobs

Jobs are stored in the database

Job Executor can be configured

−Number of Worker Threads

−Number of Jobs fetched with one database query

−Size of in-memory Queue

−Lock Time, Retry Behavior, …

Job Execution can be distributed over a Cluster

Optimizing is not a straight forward task, hard to give general advise

If you need to improve: Measure and benchmark configurations in your environment!

Job Execution

The good news: We did big performance improvements in Camunda BPM 7.2!

Improved First Level Cache (throughput increased by up to 90% if async Service Tasks are executed in a row)

Improved locking to have less Optimistic Lock Exceptions and more Jobs acquired per Acquisition. Results in bigger Clusters getting possible.

Job Execution in Camunda BPM 7.2

Recap:

Added log level “CUSTOM” for History

First Level Cache

Job Executor Acquisition Locking

Plus:

Added flush ordering (comparable to Hibernate) to minimize risk of deadlocks

Summary: Performance Improvements in 7.2

Learning #3: All performance challenges can be solved.

This is AWESOME!

Recommendation: Measure! No guessing.

camunda engine

Process Application

External Load

Generator

e.g. JMeter, HP Load Runner, CURL, …

REST

„close to production“ environment

- Measure

- JobExecutor Horizontal Scalability

- Impact of 1st level cache reuse

- Improvements Version 7.1.0 vs. Version 7.2.0

- Environment: Amazon AWS Cloud (EC2 & RDS)

Benchmark

Benchmark Setup

Client

Process Engine Node 1

Process Engine Node 2

Process Engine Node 3

Process Engine Node 4

Start Process Instance (Rest API)

Database (Postgres)

https://github.com/meyerdan/ec2-benchmark

EC2 m3.xlarge (Intel Xeon E5-2670 v2, 4 core, 15 GiB Memory)

EC2 m3.xlarge (Intel Xeon E5-2670 v2, 4 core, 15 GiB Memory)

EC2 db.m3.xlarge (Intel Xeon E5-2670 v2, 4 core, 15 GiB Memory)

Provisioned using Docker

EC2

Benchmark Setup - The process

- All service tasks „Async“ - 1st service task creates 5 variables - Variables are read by subsequent service tasks

Throughput in terms of transactions / second

No absolute Numbers

Benchmark Results

Benchmark Results (1)

Benchmark Results (1)

Benchmarks Results (2)

Benchmarks Results

Cache Off Cache On

Amazon RDS Metrics

Benchmarks Results

Cache Off

Cache On

Amazon RDS Metrics

What about true Horizontal Scalability?

What is Horizontal Scalability? Scale up the number of transactions executed by adding more processing nodes to the system. [*] [*] http://en.wikipedia.org/wiki/Scalability#Horizontal_and_vertical_scaling (Adapted)

Horizontal Scalability

transactions / sec

nodes

The current Situation

Scale number of Process Engine Nodes (JVMs) Up to a certain point

Limited possibilities for scaling the shared relational Database. In a sense this can only be scaled “up”, not “out”.

Shared Relational Database

Process Engine

Process Engine

Process Engine

Which way to go?

Distributed Datastore

Process Engine

Process Engine

Process Engine

Distributed Datastore. Use a database which is itself a distributed system and can be scaled horizontally.

- Apache Cassandra, - Apache HBase, - Distributed Caches

(Hazelcast, …) - ...

Sharding and partitioning. Distribute the state over multiple Datastores.

- Multiple instances of PostgreSQL

- Each “DB” is a Mongo DB shard

- No “DB” at all: use a filesystem journal?

- ...

Key Difference: on the right hand side, the process engine itself is “distributed” in the sense that it is aware of the distribution and sharding.

The problem with Distributed Datastores

(In the context of process engines)

1. Consistency guarantees offered by these databases (eventual consistency, ACID vs.

BASE, ...) often do not match the requirements of BPMN process execution. See:

conflicting concurrent transactions: a. Racing incoming signals (E.g.: Two Messages targeting the same event instance arrive at the

same time)

b. Joins & Synchronization (E.g.: Gateways, Multi Instance, ...)

c. Cancel Activity instance (E.g.: Interrupting Message Boundary Event)

1. Data Representation and Network Latency / Overhead: Process instance state is

composite: a. Token state / active activity instances

b. Variables

c. Task Information, …

Challenge is to find a data representation which does not lead to distribution of the state of a

single process instance across the cluster while still supporting the required access patterns.

2. Significant differences between individual technologies while there are no

industry standards in place yet. (Different with SQL).

Sharding => Distributed yet Local

Scale horizontally...

Each “shard / node” maitains its state locally Partitioning workflow instance state - Each process instance lives inside a single shard / partition => local data consistency easy to guarantee, => easy to access efficiently => Support range of different persistence engines (Relational Database, Non-Relational Databases, …)

Process

Engine

Flexible Architecture

...

Reality @ zalando 2014

Process

Engine

Process

Engine

The simplest case A single process engine node

running on top of a conventional database.

A medium Scenario Horizontally scale on top of a

conventional database.

Massive Compute Cluster 500 Nodes ?

All of this should be possible with one unified architecture!

No more Search!

The catch

“Find Process Instance for order with ID 43543242” ??

???

Human Workflow (Build Task Lists) History: Monitoring, Reporting, … Message Correlation

When is „Search“ required?

Message Correlation

The Problem to solve

Workflow Instance State for order with ID 435345

Incoming Message: “customer cancelled Order

with ID 435345”

Yes, but for non-workflow execution Use Cases

Use Search Index?

(A)sync Updates

Search Index (Near Realtime)

Tasklist Queries, Monitoring,...

Vision

History Tasks Core Process Execution

Signal / Cancel Activity Instance by Id Correlate Message

Query for List of Tasks

Monitoring, Reports

Real Time, Strongly Consistent Horizontally scalable through sharding

Multiple persistence technologies possible

Near Real Time, Eventually Consistent Use best technology for the Job.

Async Event Stream

But still...

History Tasks Core Process Execution

Signal / Cancel Activity Instance by Id Correlate Message

Query for List of Tasks

Monitoring, Reports

In the simplest case!

Learning #4: You can do true horizontal clustering with the engine which exists today! There is no need for No-SQL persistence in the core engine.

Learning #5: Camunda is really damn smart :-)

Camunda BPM Performance is already awesome

However: We are continuously improving performance

There are strategies to solve specific performance challenges

There is no limit in scalability

Summary

Start now!

Open Source Edition • Download:

www.camunda.org • Docs, Tutorials etc. • Forum • Meetings

Enterprise Edition • Trial:

www.camunda.com • Additional Features • Support, Patches etc. • Consulting, Training

http://camunda.com/bpm/consultation/

[email protected] | US +1.415.800.3908 | DE +49 30 664040 900

Thank you!

Questions?