NOTICE: SUPPORT FOR THIS PROJECT HAS ENDED
This projected was owned and maintained by Walmart. This project has reached its end of life and Walmart no longer supports this project.
We will no longer be monitoring the issues for this project or reviewing pull requests. You are free to continue using this project under the license terms or forks of this project at your own risk. This project is no longer subject to Walmart's bug bounty program or other security monitoring.
We recommend you take the following action:
- Review any configuration files used for build automation and make appropriate updates to remove or replace this project
- Notify other members of your team and/or organization of this change
- Notify your security team to help you evaluate alternative options
For security reasons, Walmart does not transfer the ownership of our primary repos on Github or other platforms to other individuals/organizations. Further, we do not transfer ownership of packages for public package management systems.
If you would like to fork this package and continue development, you should choose a new name for the project and create your own packages, build automation, etc.
Please review the licensing terms of this project, which continue to be in effect even after decommission.
BigBen
is a generic, multi-tenant, time-based event scheduler and cron
scheduling framework based on Cassandra
and Hazelcast
It has following features:
- Distributed -
BigBen
uses a distributed design and can be deployed on 10's or 100's of machines and can be dc-local or cross-dc - Horizontally scalable -
BigBen
scales linearly with the number of machines. - Fault tolerant -
BigBen
employs a number of failure protection modes and can withstand arbitrary prolonged down times - Performant -
BigBen
can easily scale to 10,000's or even millions's of event triggers with a very small cluster of machines. It can also easily manage million's of crons running in a distributed manner - Highly Available - As long as a single machine is available in the cluster,
BigBen
will guarantee the execution of events (albeit with a lower throughput) - Extremely consistent -
BigBen
employs a single master design (the master itself is highly available withn-1
masters on standby in ann
cluster machine) to ensure that no two nodes fire the same event or execute the same cron. - NoSql based -
BigBen
comes with default implementation withCassandra
but can be easily extended to support otherNoSql
or evenRDBMS
data stores - Auditable -
BigBen
keeps a track of all the events fired and crons executed with a configurable retention - Portable, cloud friendly -
BigBen
comes as application bundled aswar
or an embedded lib asjar
, and can be deployed on any cloud,on-prem
orpublic
BigBen
can be used for a variety of time based workloads, both single trigger based or repeating crons.
Some of the use cases can be
- Delayed execution - E.g. if a job is to be executed 30 mins from now
- System retries - E.g. if a service A wants to call service B and service B is down at the moment, then service A can schedule an exponential backoff retry strategy with retry intervals of 1 min, 10 mins, 1 hour, 12 hours, and so on.
- Timeout tickers - E.g. if service A sends a message to service B via
Kafka
and expects a response in 1 min, then it can schedule atimeout check
event to be executed after 1 min - Polling services - E.g. if service A wants to poll service B at some frequency, it can schedule a cron to be executed at some specified frequency
- Notification Engine -
BigBen
can be used to implementnotification engine
with scheduled deliveries, scheduled polls, etc - Workflow state machine -
BigBen
can be used to implement a distributedworkflow
with state suspensions, alerts and monitoring of those suspensions.
BigBen
was designed to achieve the following goals:
- Uniformly distributed storage model
- Resilient to hot spotting due to sudden surge in traffic
- Uniform execution load profile in the cluster
- Ensure that all nodes have similar load profiles to minimize misfires
- Linear Horizontal Scaling
- Lock-free execution
- Avoid resource contentions
- Plugin based architecture to support variety of data bases like
Cassandra, Couchbase, Solr Cloud, Redis, RDBMS
, etc - Low maintenance, elastic scaling
See the blog published at Medium
for a full description of various design elements of BigBen
BigBen
can receive events in two modes:
- kafka - inbound and outbound Kafka topics to consume event requests and publish event triggers
- http - HTTP APIs to send event requests and HTTP APIs to receive event triggers.
It is strongly recommended to use kafka
for better scalability
Request and Response channels can be mixed. For example, the event requests can be sent through HTTP APIs but the event triggers (response) can be received through a Kafka Topic.
BigBen
has a robust event processing guarantees to survive various failures.
However, event-processing
is not same as event-acknowledgement
.
BigBen
works in a no-acknowledgement mode (at least for now).
Once an event is triggered, it is either published to Kafka
or
sent through an HTTP API
. Once the Kafka
producer returns success, or HTTP API
returns non-500 status code,
the event is assumed to be processed and marked as such in the system.
However, for whatever reason if the event was not processed and resulted in an error
(e.g. Kafka
producer timing out, or HTTP API
throwing 503
),
then the event will be retried multiple times as per the strategies discussed below
Multiple scenarios can cause BigBen
to be not able to trigger an event on time. Such scenarios are called
misfires. Some of them are:
-
BigBen
's internal components are down during event trigger. E.g.BigBen
's data store is down and events could not be fetchedVMs
are down
-
Kafka
Producer could not publish due to loss of partitions / brokers or any other reasons -
HTTP API
returned a 500 error code -
Any other unexpected failure
In any of these cases, the event is first retried in memory using an exponential back-off strategy.
Following parameters control the retry behavior:
- event.processor.max.retries - how many in-memory retries will be made before declaring the event as error, default is 3
- event.processor.initial.delay - how long in seconds the system should wait before kicking in the retry, default is 1 second
- event.processor.backoff.multiplier - the back off multiplier factor, default is 2. E.g. the intervals would be 1 second, 2 seconds, 4 seconds.
If the event still is not processed, then the event is marked as ERROR
.
All the events marked ERROR
are retried up to a configured limit called events.backlog.check.limit
.
This value can be an arbitrary amount of time, e.g. 1 day, 1 week, or even 1 year. E.g. if the the limit
is set at 1 week
then any event failures will be retried for 1 week
after which, they will be permanently
marked as ERROR
and ignored. The events.backlog.check.limit
can be changed at any time by changing the
value in bigben.yaml
file and bouncing the servers.
BigBen
shards events by minutes. However, since it's not known in advance how many events will be
scheduled in a given minute, the buckets are further sharded by a pre defined shard size. The shard size is a
design choice that needs to be made before deployment. Currently, it's not possible to
change the shard size once defined.
An undersized shard value has minimal performance impact, however an oversized shard value may
keep some machines idling. The default value of 1000
is good enough for most practical purposes as long as
number of events to be scheduled per minute exceed 1000 x n
, where n
is the number of machines in the cluster.
If the events to be scheduled are much less than 1000
then a smaller shard size may be chosen.
Each bucket with all its shards is distributed across the cluster for execution with an algorithm that ensures a
random and uniform distribution. The following diagram shows the execution flow.
Multiple tenants can use BigBen
in parallel. Each one can configure how the events will be delivered once triggered.
Tenant 1 can configure the events to be delivered in kafka
topic t1
, where as tenant 2 can have them delivered
via a specific http
url. The usage of tenants will become more clearer with the below explanation of BigBen
APIs
BigBen is dockerized and image (bigben
) is available on docker hub. The code also contains
scripts, which start cassandra
, hazelcast
and app
.
To quickly set up the application for local dev testing, do the following steps:
git clone $repo
cd bigben/build/docker
- execute
./docker_build.sh
- start cassandra container by executing
./cassandra_run.sh
- start app by executing
./app_run.sh
- To run multiple app nodes
export NUM_INSTANCES=3 && ./app_run.sh
- wait for application to start on port
8080
- verify that
curl http://localhost:8080/ping
returns200
- Use
./cleanup.sh
to stop and remove allBigBen
related containers
BigBen
can be run without docker as well. Following are the steps
git clone $repo
cd bigben/build/exec
- execute
./build.sh
- execute
./app_run.sh
You can set the following environment properties
APP_CONTAINER_NAME
(default bigben_app)SERVER_PORT
(default 8080)HZ_PORT
(default 5701)NUM_INSTANCES
(default 1)LOGS_DIR
(default bigben/../bigben_logs)CASSANDRA_SEED_IPS
(default $HOST_IP)HZ_MEMBER_IPS
(default $HOST_IP)JAVA_OPTS
#How to override default config values?
BigBen
employs an extensive override system to allow someone to override
the default properties. The order of priority is system properties > system env variables >
overrides > defaults
The overrides can be defined in config/overrides.yaml
file.
The log4j.xml
can also be changed to change log behavior without
recompiling binaries
Following are the steps to set up Cassandra
:
- git clone the
master
branch - Set up a Cassandra cluster
- create a keyspace
bigben
inCassandra
cluster with desired replication - Open the file
bigben-schema.cql
and executecqlsh -f bigben-schema.cql
GET /events/cluster
- response sample (a 3 node cluster running on single machine and three different ports (5701, 5702, 5703)):
{
"[127.0.0.1]:5702": "Master",
"[127.0.0.1]:5701": "Slave",
"[127.0.0.1]:5703": "Slave"
}
The node marked Master
is the master node that does the scheduling.
A tenant can be registered by calling the following API
POST /events/tenant/register
- payload schema
{
"$schema": "http://json-schema.org/draft-04/schema#",
"type": "object",
"properties": {
"tenant": {
"type": "string"
},
"type": {
"type": "string"
},
"props": {
"type": "object"
}
},
"required": [
"tenant",
"type",
"props"
]
}
-
tenant
- specifies a tenant and can be any arbitrary value. -
type
- specifies the type oftenant
. One of the three types can be used- MESSAGING - specifies that
tenant
wants events delivered via a messaging queue. Currently,kafka
is the only supported messaging system. - HTTP - specifies that
tenant
wants events delivered via an http callback URL. - CUSTOM_CLASS - specifies a custom event processor implemented for custom processing of events
- MESSAGING - specifies that
-
props
- A bag of properties needed for each type of tenant. -
kafka sample:
{
"tenant": "TenantA/ProgramB/EnvC",
"type": "MESSAGING",
"props": {
"topic": "some topic name",
"bootstrap.servers": "node1:9092,node2:9092"
}
}
- http sample
{
"tenant": "TenantB/ProgramB/EnvC",
"type": "HTTP",
"props": {
"url": "http://someurl",
"headers": {
"header1": "value1",
"header2": "value2"
}
}
}
GET /events/tenants
POST /events/schedule
Payload - List<EventRequest>
EventRequest
schema:
{
"$schema": "http://json-schema.org/draft-04/schema#",
"type": "object",
"properties": {
"id": {
"type": "string"
},
"eventTime": {
"type": "string",
"description": "An ISO-8601 formatted timestamp e.g. 2018-01-31T04:00.00Z"
},
"tenant": {
"type": "string"
},
"payload": {
"type": "string",
"description": "an optional event payload, must NOT be null with deliveryOption = PAYLOAD_ONLY"
},
"mode": {
"type": "string",
"enum": ["UPSERT", "REMOVE"],
"default": "UPSERT",
"description": "Use REMOVE to delete an event, UPSERT to add/update an event"
},
"deliveryOption": {
"type": "string",
"enum": ["FULL_EVENT", "PAYLOAD_ONLY"],
"default": "FULL_EVENT",
"description": "Use FULL_EVENT to have full event delivered via kafka/http, PAYLOAD_ONLY to have only the payload delivered"
}
},
"required": [
"id",
"eventTime",
"tenant"
]
}
GET /events/find?id=?&tenant=?
POST /events/dryrun?id=?&tenant=?
fires an event without changing its final status
coming up...