Link Header-based Invalidation of Caches
- 1. Using HTTP Link: Header for Gateway Cache Invalidation Mike Kelly <mike@mykanjo.co.uk> Michael Hausenblas <michael.hausenblas@deri.org>
- 2. What is a gateway cache? "reverse proxy cache" A layer between all clients and destination server Objective: Minimize demand on destination server Not so concerned with reducing bandwith
- 3. How do they work? They can leverage the 3 principal caching mechanisms: • Expiration • Validation • Invalidation HTTP has mechanisms for each of these
- 4. Expiration-based caching < 200 OK < Content-Type: text/html < Cache-Control: public, s-maxage=600 < .... Pros: + Simple + No contact with server until expiration Cons: - Inefficient - Difficult to manage
- 5. Validation-based caching < 200 OK < ETag: "686897696a7c876b7e" > GET /example > If-None-Match: "686897696a7c876b7e" < 304 Not Modified Pros: + Reduces bandwidth + Ensures freshness Cons: - Server handling every request - Generating 304 still costs processing and I/O
- 6. Expiration+Validation caching < 200 OK < ETag: "686897696a7c876b7e" < Cache-Control: public, s-maxage=600 Pros: + Expiration reduces contact with server + Validation reduces bandwidth Cons: - "Worst case" inefficiency - Still managing caching rules
- 7. Invalidation-based caching - Responses fresh until invalidated (by non-safe requests) In HTTP: PUT POST PATCH DELETE (PURGE?)
- 8. How is this possible? Product of adhering to constraints of REST, particularly: Uniform Interface + Self-descriptive messages Intermediaries can make assertions about client-server interactions.
- 9. Invalidation-based caching Pros: + Caches have self-control + "Best case" efficiency + Ensured freshness* Cons: - Only reliable for gateway caches - Impractical* * (sort of)
- 10. Cache invalidation in practice Two main problems for cache invalidation arise from pragmatism and trade-offs in resource granularity and identification: • The "Composite Problem" • The "Split-resource Problem"
- 11. Composite Problem Perfect World: <collection> <item rel="item" href="/items/123" /> <item rel="item" href="/items/asdf" /> <item rel="item" href="/items/foobar" /> </collection> Real World: <collection> <item rel="item" href="/items/123"> <title>Item 123</title> <content>Content for item 123 - an example of embedded state</content> </item> <item rel="item" href="/items/asdf"> <title>Item asdf</title> <content>This state is also embedded</content> </item> <item rel="item" href="/items/foobar"> <title>FooBar</title> <content>Yet more embedded state!! :(</content> </item> </collection>
- 12. Composite Problem What effect would the following interaction have on the composite collection it belongs to? > PUT /composite-collection/item123 < 200 OK
- 13. The Split-resource Problem Given /document resource with representations: /document.html /document.xml /document.json When a client does this: PUT /document Then invalidation of each representation is invisible to intermediaries
- 15. .. The Solution Beef up the uniform interface: Express these common types of resource dependency as control data using Link header and standard link relations This increases: - Self-descriptiveness of messages - Visibility "Link Header-based Invalidation of Caches" (LHIC)
- 16. LHIC-I Express dependency in response to an invalidating request > PUT /composite-collection/item123 < 200 OK < Link: </composite-collection>; < rel="http://example.org/rels/dependant"
- 17. LHIC-II Express dependencies in initial cacheable responses > GET /document.html < 200 OK < Link: </document>; < rel="http://example.org/rels/dependsOn" > GET /document.xml < 200 OK < Link: </document>; < rel="http://example.org/rels/dependsOn" > GET /document.json < 200 OK < Link: </document>; < rel="http://example.org/rels/dependsOn" > PUT /document < 200 OK
- 18. Comparison LHIC-I + More dynamic control of invalidation - DoS risk - Invalidation does not cascade LHIC-II + No DoS risk + Cascading invalidation - Complexity
- 19. Conclusion LHIC injects lost visibility. Resulting mechanism: + Very efficient + Ensures freshness + Easily managed + Leverages existing specs - Only for gateway caching + Combine Invalidation (gateway) & Validation (client)
- 20. Considerations Resource state altered outside of uniform interface - Don't do that - Reintroduce expiration and validation Peering - Further research Size limits for HTTP headers