BigBen

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 with n-1 masters on standby in an n cluster machine) to ensure that no two nodes fire the same event or execute the same cron.
• NoSql based - BigBen comes with default implementation with Cassandra but can be easily extended to support other NoSql or even RDBMS 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 as war or an embedded lib as jar, and can be deployed on any cloud, on-prem or public

Use cases

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 a timeout 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 implement notification engine with scheduled deliveries, scheduled polls, etc
• Workflow state machine - BigBen can be used to implement a distributed workflow with state suspensions, alerts and monitoring of those suspensions.

Architectural Goals

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

Design and architecture

See the blog published at Medium for a full description of various design elements of BigBen

Events Inflow

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 and HTTP APIs to receive event triggers.

It is strongly recommended to use kafka for better scalability

Event Inflow diagram

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.

Event processing guarantees

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

Event misfire strategy

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 fetched
  • VMs 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.

Event bucketing and shard size

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.

Multi shard parallel processing

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.

Multi-tenancy

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

How to deploy BigBen?

Following are the steps to set up BigBen:

1. git clone the master branch
2. Set up a Cassandra cluster
3. create a keyspace bigben in Cassandra
4. Open the file bigben-schema.cql and execute the commands on the cql prompt
5. Open the file bigben.yaml
   • Modify the cassandra related properties
   • Modify the hazelcast related properties
     • add the comma separated IPs of nodes participating in the cluster
     • modify any other property if needed
   • Uncomment the processor.class property if you plan to use kafka for receiving events
     • Go to kafka section and add the appropriate topic name
     • Add / Modify any other properties as needed
   • Check the buckets section and set the backlog.check.limit to some value (e.g. set 1440 for one day backlog check limit)
   • If you don't want to use cron module, remove it from under the modules section
6. run mvn clean install -DskipTests
7. copy the app/target/bigben.war to the server location in all the nodes
8. start the server
9. wait for sometime, then call GET /events/cluster API and see that a master is chosen

APIs

cluster

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.

tenant registration

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 of tenant. 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

• 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"
          }
     }
}

fetch all tenants:

GET /events/tenants

event scheduling

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"
    },
    "mode": {
      "type": "string",
      "enum": ["UPSERT", "REMOVE"]
    }
  },
  "required": [
    "id",
    "eventTime",
    "tenant"
  ]
}

find an event

GET /events/find?id=?&tenant=?

dry run

POST /events/dryrun?id=?&tenant=?

fires an event without changing its final status

cron APIs

coming up...