hadoop-common-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Apache Wiki <wikidi...@apache.org>
Subject [Hadoop Wiki] Update of "HedWig" by mahadevkonar
Date Wed, 17 Nov 2010 22:26:52 GMT
Dear Wiki user,

You have subscribed to a wiki page or wiki category on "Hadoop Wiki" for change notification.

The "HedWig" page has been changed by mahadevkonar.


New page:
== Hedwig ==

Hedwig is a _publish-subscribe_ system designed to carry large amounts of data across the
internet in a *guaranteed-delivery* fashion from those who produce it (_publishers_) to those
who are interested in it (_subscribers_). The goals of Hedwig are:

   1 *Guaranteed Delivery*: The messages may not be lost even when a particular subscriber
goes offline for extended periods of time.
   1 *Topic-based*: Publishes and subscribes happen to a topic. The system should scale to
~10^6 topics with ~10 subscribers on each topic.
   1 *Incremental Scalability*: It should be possible to scale capacity by adding servers
on the fly. The up-front hardware investment should not be huge.
   1 *High availability*: The system should remain available in the presence of single server
failure without manual intervention. 

While there are a lot of commercial and open-source products in the general pub-sub category,
none of them satisfy all the above 4 requirements (they mostly fail on 3 and 4).

*Hedwig is in the early stages of development and not generally available right now*.


<img src="%ATTACHURLPATH%/hedwig.jpg" height="200"/>


---+++ Architecture

At the highest level, Hedwig is a collection of *regions* spread across the internet. Any
region may publish on a topic, and those messages must be delivered to any subscriber in any
region that has subscribed to that topic.

<img src="%ATTACHURLPATH%/global.png" width="500"/>


Now digging down into a region, it consists of a collection of *hub servers*. Hub servers
aggregate messages published in a region and persist them. They also subscribe to hubs in
other regions to listen for messages that their clients are subscribed to. Clients always
subscribe only to  local hub servers. Hedwig plans to use [[http://hadoop.apache.org/zookeeper/][Zookeeper]]
for persistence of metadata, and [[https://issues.apache.org/jira/browse/ZOOKEEPER-276][Bookkeeper]]
for persistence of actual messages.

<img src="%ATTACHURLPATH%/region.png" width="500"/>

Topics are randomly split over hubs. When the hub responsible for a topic fails, another hub
should take over responsibility of the topic.


Now digging into a hub, it consists of 6 components:

<img src="%ATTACHURLPATH%/hub.png" width="500"/>

   1 *Network-I/O component*: For high throughput, we use Java NIO through a framework called
   1 *Topic Manager*: Maintains and coordinates ownership of topics among hubs. It is responsible
for doing automatic failover when a hub dies. It will be a Zookeeper client.
   1 *Subscription Manager*: Maintains information about which subscriptions exist in the
system, and their _consume-points_: the point in the topic until which a subscriber has already
received and acknowledged, so that the next delivery of a message can start from this point.
   1 *Persistence Manager*: Persists messages in a reliable way so that they can be retrieved
sequentially later. It will be a Bookkeeper client.
   1 *Remote Subscriber*: Subscribes to hubs in other regions for the topics, so that the
messages published there can also reach the local clients. This component will be our own
Hedwig Java client.
   1 *Delivery Manager*: This component will be responsible for delivering messages to the
subscribers. The subscribers can be either our own local clients or the hub in other regions.

---+++ Code-Level Design & Details

The hedwig project consists of 3 modules:

   * [[http://svn.corp.yahoo.com/view/yahoo/yrl/hedwig/trunk/hedwig/protocol/src/main/protobuf/PubSubProtocol.proto?view=markup][Protocol]]:
This is a simple module that specifies the client-server protocol used by Hedwig. The protocol
is specified as a [[http://code.google.com/p/protobuf/][Protocol Buffers]] file. Protocol
buffers can generate serialization and deserialization code for us in multiple languages.
   * [[HedwigClient][Client]]: We support both a c++ and a java client library. The client
module obviously depends on the protocol module. 
   * [[HedwigServer][Server]]: This the major chunk of the system. A server is responsible
for certain topics, and accepting publish and subscribe requests for them. The server uses
(and hence depends on) the hedwig java client to subscribe to topics in other regions.   

Get started by checking out the repository:

Follow the directions in the BUILD.txt

The following *coding conventions* are to be followed:
   1 Indentation only with 4 spaces, no tabs (adjustable using Eclipse settings).
   1 Use curly braces even for single-line ifs and elses. I am sure you have encountered at
least 1 bug in your life because the else got associated with the wrong if.
   1 No System.outs (only logging)
   1 Javadoc (even if brief) for every class. No author tags.
   1 We are building a high-throughput server. Watch out for the following performance crippling
      * If you are logging something that requires string concatenation, even if it is at
debug level, you pay the price of concatenation. Hence surround by =if(isDebugEnabled())=
      * Use !StringBuilder while building long strings, don't just use +'s.
   1 Remember that we are multithreaded which is a double-edged sword.
      * Protect your data structures where needed with a lock.
      * However locking needs to be well-understood and documented. Please bring it up in
the design discussions. 

---+++ Test

Every class that has a non-abstract method must have a test case. Maven directory structure
is already set up to be able to write and run tests easily.
---+++ Current Tasks

---++++ Protocol (Ben)

   1 use region name as subscriberID for hub subscribers. This needs only an explanatory comment
   1 Put a transaction-id in publishes. This can help us in firing off multiple publishes
at a time.
   1 Put a version string in every request. 
   1 Figure out our authentication and access control story. It's important to have one.

---++++ Server

   1 Design a push interface for reading rather than the current pull-based (Utkarsh)
   1 How do topics get created? Is it on the first subscription? Who gets the ownership for
the topic, and how do we ensure that this is load balanced across the hubs? (Brian)
   1 Figure out a rich admin interface and how to implement it with the least effort - Possibly
jmx (Erwin)
   1 Cross-region subscription problem (and hence log collection problem): How do we avoid
making the cross-region call when we receive a local subscription for a brand new topic. If
we don't make the cross-region call, then log collection must be performed only after all
regions agree that no one is interested. (Adam)
   1 Designing the delivery manager, especially the management of readahead of messages for
subscribers, and caching the recently published messages. While designing this interface,
keep it pluggable to different policies. 
   1 Consider whether bookkeeper client library needs to be written using asynchronous network
I/O. IOW, is 1 thread per "bookie" acceptable?

---+++ Notes

   * [[HedwigTopicManagement][Topic management]]

View raw message