Scale horizontally (scale out): Separate system tiers on different environments: database, Solr, memcached, push, web/app servers. Separation helps in scaling up/down tiers individually.
Don’t use filesystems for storage, unless it is a distributed filesystem (AWS S3)
Don’t involve your app server in long requests/responses. Slow clients may block your server and cause longer request queues (depends on implementation).
If you want to receive an upload get it through S3 with some work on the client side.
If you want to send a huge response, either stream it using a streaming capability of your app server, or generate it using a background job that stores it finally on S3 and sends the direct link when done through the app using push or through email.
Don’t clutter your app server memory with language bindings, use Apache Thrift or Google Protocol Buffers to communicate between different environments
Use Apache Solr (Lucene over HTTP) to query large data even if you don’t have full text search, it can be used for scoping and faceting as well (think SQL WHERE and GROUP BY)
Autoscale your web/app servers depending on traffic. Monitoring shows you traffic metrics and HireFire will autoscale your heroku dynos
Use clients-side rendering (Javascript templates) to get rid of the rendering time on the server
Use caching in different layers (memcached) to be nice on your servers
Asset hosting, static assets (javascripts, stylesheets, html templates, images, …) have nothing to do with your app servers so host them somewhere else (CloudFront/S3)