Phew

I have been coding in JSP for a while now and worked on the Struts framework as well. The servlet API's allow adding of instances to various levels of scope within the webapp. Like Application scope for the entire webapp, Session scope for the user session, Request scope for a particular request. Request scope is useful as the particular request response can be generated in parts by chaining servlets together. Now these methods make life pretty easy. Something which will used throughout the webapp can be kept at application scope and similarly at varying degrees of granularity at other scope. Like on correct credentials for a login form, a String can be stored in session scope.

This would always disturb me. Simply because client specific data was being stored in-memory. Upto how many users can be supported in such an architecture? even if 10 bytes are stored per client, that can count to a lot of memory for say 10k users. Then the overhead of caching, maintaining state of live data and associating it with the client.

Initially storing stuff in scope would make me think of the size of the objects I used to store. Then I realised frameworks like Struts store huge amounts of data at various scope. Like the action form etc in the request scope. So that made me feel that this is not such an issue. If such relatively heavy objects can be stored then that must be fine. And now with JSF etc these object are just getting larger.


Then I read on the Php Share-Nothing architecture. So the Share-Nothing architecture advocates... simply sharing nothing about the client on the server. As simple as that. Instead of storing stuff in memory, store details in a DB and make DB calls always. Now I am not actually sure how performant this is. But DB's have been around for a long while and have been well tuned. Plus DB's can be easily made to work in parallel redundant mode and have very good features.

Again I know I am being premature to be so anti-in-memory stuff. Many large systems have been built, especially in the enterprise. But it just does not feel right. In the end it seems easier to scale horizontally with the Share-Nothing approach. The in-memory approach seems to force towards vertical scaling.

Some examples... Flickr and Yahoo are two php based webapps. And they dont get bigger than them. Ebay is the biggest java based webapp i know of. But Google for ebay architecture and they too have a custom Share-Nothing like system in place. Gmail too uses Java, but have some super optimizations in place.

Rails is one very hot webapp framework right now. This is one of the more intelligent discussions I have read. Coupled with this blog post Rails seems like something I would like to learn soon. These guys seem to be aware of both php and Java webapp dev. Any idea on their principles on this issue?

What dyu guys think? Any of you really stick to the Share-Nothing principle?