Difference between revisions of "Zaqar/specs/havana"
Revision as of 17:04, 19 June 2013
Marconi: Cloud Message Queuing for OpenStack
This specification formalizes the requirements and design considerations captured during one of the Grizzly Summit working sessions to initiate a message bus project for OpenStack. As the project evolves, so too will its requirements, so this specification is only meant as a starting point.
Here's a brief summary of how Marconi works:
- Clients post messages via HTTP to Marconi. The URL contains a tenant ID.
- Marconi persists messages according to either a default TTL, or one specified by the client.
- Clients poll Marconi for messages.
- Clients may optionally claim a batch of messages, hiding them from other clients. Once the client has processed each message, it can delete it from the server. In this way, Marconi provides a mechanism for ensuring each message is processed once and only once.
The lack of an integrated cloud message bus service is a major inhibitor to OpenStack adoption. While Amazon has SQS and SNS, OpenStack currently provides no alternatives.
OpenStack needs a multi-tenant message bus that is fast, efficient, durable, horizontally-scalable and reliable.
The Marconi project will address these needs, acting as a compliment to the existing RPC infrastructure within OpenStack, while providing multi-tenant services that can be exposed to applications running on public and private clouds.
1. Distribute tasks among multiple workers (transactional job queues)
2. Forward events to data collectors (transactional event queues)
3. Publish events to any number of subscribers (pub-sub)
4. Send commands to one or more agents (RPC via point-to-point or pub-sub)
5. Request information from an agent (RPC via point-to-point)
6. Monitor a Marconi deployment (DevOps)
Marconi's design philosophy is derived from Donald A. Norman's work regarding The Design of Everyday Things:
The value of a well-designed object is when it has such a rich set of affordances that the people who use it can do things with it that the designer never imagined.
Goals related to the above:
- Emergent functionality, utility
- Modular, pluggable code base
- REST architectural style
Principles to live by:
- Versioned API
- Implemented in Python, following PEP 8 and pythonic idioms
- Modular, driver-based architecture
- Async I/O
- Low response time, turning around requests in 20-50ms (or better), even under load
- High throughput, serving millions of reqs/min with a small cluster
- Thousands of req/sec per queue (?)
- 100's of thousands of queues per tenant
- Horizontal scaling of both reads and writes
- Support for HA deployments
- Guaranteed delivery
- Best-effort message ordering
- Server generates all IDs
- Gzip'd large messages
- LZ4 compression for messages at rest
- Secure (audited code, end-to-end HTTPS support, penetration testing, etc.)
- Schema validation
- Auth caching
- Multi-Transport (Http, ZMQ)
- Eventing and work queuing semantics
- Opaque payload (although must be valid JSON)
- Max payload size of 64K
- Batch message posting and querying
- Keystone auth driver (service catalog may return endpoints for different regions and/or different characteristics)
- REPL for debugging, testing, diagnostics
- Client libraries for Python, PHP, Java, and C#
- Auto-generated audit river (read-only queue) for actions and state changes
- Delayed delivery (UNDER DISCUSSION - MAY OR MAY NOT BE IN 1.0)
TODO: Create blueprints for these, prioritize
Brainstormed features, listed in no particular order:
- PATCH support for updating queue metadata
- Set/get arbitrary queue metadata
- Kombu Integration
- API tokens tied to a specific app and a specific queue, OAuth?
- Message signing
- Delayed messages
- Standalone control panel or at least a simple admin/dashboard app for ops
- JSON-P support (may need to use the while(1); trick to prevent XSS attacks)
- Multi-get (specify a list of queues to query in a single request)
- Tag-based filtering
- Includes a way to return in one call, everything with or without the tag (OR semantics) to afford fanout.
- XML support
- LZ4 body compression (in WSGI server as well as client libs)
- Response caching
- Authorization (based on tags and/or queues)
- Cross-tenant sharing (need to define business case)
- Temporal queries
- Ruby client library
- PHP client library
- Cross-regional replication
- Horizon plug-in
- Ceilometer data provider
- PyPy support
- HTTP 2.0 support
Marconi may be used to support other services that provide the following functionality, but will not embed these abilities directly within its code base.
- Any kind of push notifications over persistent connections (leads to complicated state management and poor hardware utilization)
- Forwarding notifications to email, SMS, Twitter, etc. (ala SNS)
- Forwarding notifications to web hooks
- Forwarding notifications to APNS, GCM, etc.
- Scheduling-as-a-service (ala IronWorker)
- Metering and monitoring solutions
Marconi will use a micro-kernel architecture. Auth, transport, storage, cache, logging, monitoring, etc. will all be implemented as drivers or exposed with standard protocols, allowing vendors to customize Marconi to suit.
Endpoint controllers define the interface between storage and transport. More info.
Possible frameworks that can help realize a highly modular design:
- WSGI transport using Falcon
- Endpoints: HTTP(S), ZMQ
- Auth: Keystone middleware
- Storage: MongoDB
- Logging: Standard library logging
- Monitoring: Statsd, as well as HTTP stats page?
- Self-host via gevent.http or ZMQ
- Host with a WSGI server.
- Requires writing a small bootstrap script to load the kernel and export the app callable.
- Bootstrap script also allows full programmatic customization of logging
See the Marconi API spec. [ROUGH DRAFT]
All development will be done TDD-style using nose and testtools. Pair programming may happen on accident (or even on purpose). Eventually we'll add integration, performance, and security tests, and get everything automated in a nice and tidy CI pipeline.