Medium-long term roadmap - 2011/03/01 Legend: '+' = done, '-' = todo, '*' = done except doc 1.5 (ETA 2010/12/31) : - count number of monitor requests on frontends, that's the only way to explain the possible huge difference between frontend and backend sessions. - ACL to report number of used entries in a table - server-side HTTP keepalive => maybe with limitation to only reuse connections that don't depend on layer7 in a first time (just check the target). - POST parameter extraction and size/speed measurement to use in ACLs - return-html code xxx [ file "xxx" | text "xxx" ] if - return-raw [ file "xxx" | text "xxx" ] if - avg connect time, response time, connect errors, response errors in stats - add a last activity date for each server (req/resp) that will be displayed in the stats. It will be useful with soft stop. - add the ability to only dump response errors to more easily detect anomalies without being polluted with attacks in requests. - add support for server-side unix sockets - have multi-criteria analysers which subscribe to req flags, rsp flags, and stream interface changes. This would result in a single analyser to wait for the end of data transfer in HTTP. - implement support for "connection freeze" after accept. A list of frozen connections should be maintained so that it is possible to recycle them when new file descriptors are required. - support for time-ordered priority queues with ability to add an offset based on request matching. Each session will have one ebtree node to be attached to whatever queue the session is waiting in. - assign a nice priority based on ACLs. - dontlog if (front/back) - fix "PR--" flags when accessing stats - pattern extraction is needed for ACLs and stickiness. It would work like this : acl [-i] ... All ACL fetch method currently available would be transformed into pattern extraction methods. That way we could stick on hdr(x-forwarded-for) or use source 0.0.0.0 usesrc (such as "hdr_ip(headername)"). Note that ACLs sometimes need iterative matching/extraction. - add support for complex pattern extraction rules : pattern = | '{' pattern_expr '}' pattern_expr = [ ... ] - support loading data sets from files + present/not present (eg: netmasks) - pattern conversion per prefixes. Eg: convert src IP to country. - automatically compute fullconn for backends : by default, set it to 10% of the sum of the maxconn of all unique frontends which reference it via use_backend, default_backend or that are in the same listen. - what to do with data after a POST and how to detect some data were received when responding ? In theory we should read everything because the TCP stack does not notify us that the FIN was acked. In practice, reading just before closing should be enough. Right now we simply read whatever comes after the POST. => switch the connection to a "drain" state, where it monitors its output queue on each I/O and where it can be stolen if fds are missing. - half-closed timeouts ? - add a flag in logs to indicate keep-alive requests ? - make it possible to condition a timeout on an ACL - forwardfor/originalto except with IPv6 - have a callback function which would be called after a server is selected, for header post-processing. That would be mainly used to remove then add the server's name or cookie in a header so that the server knows it. - remove lots of remaining Alert() calls or ensure that they forward to send_log() after the fork. - initcwnd parameter for bind sockets DONE: * rename L4 acls as L6 ACLs when some content is involved * add new L4 ACL checks immediately after accept, before even allocating the buffers ("connection {accept|reject|delay|freeze} {if|unless}"). * make new patterns available based on stickiness matching : * number of entries in table for the matched pattern * same after having increased the match counter * add support for concurrency match in tables * just like stickiness, but counted per session (or request), increased on first match and decreased at end of request or connection. This requires that the session has a list of matched terms that must be released at the end. * http_req_first ACL * expirable cookies + "preserve" * ECV, LDAPv3 & MySQL checks * configurable check buffer size * stats + ON/OFF * halog: sort by URL * "PROXY" protocol * add support for client-side unix sockets * hash: rehash non-consistent hashes with chash() for more randomness. * add an error ID in captures to ease new error detection for scripts. * try to remove srv==NULL internally and assign a dummy server to each backend for dispatch, http_proxy and transparent modes. => done differently with the target descriptors. The dummy server code exists in the "dummysrv" branch which will die since it does not make sense anymore. 1.6 (will probably change anyway) : - wait on resource (mem, socket, server's conn, server's rate, ...) - bandwidth limits - create internal services and make stats, CLI, etc... part of that. - use_server ... if ... - buddy servers to build defined lists of failovers. Detect loops during the config check. server XXX buddy YYY server YYY # may replace XXX when XXX fails - spare servers : servers which are used in LB only when a minimum farm weight threshold is not satisfied anymore. Useful for inter-site LB with local pref by default. - add support for event-triggered epoll, and maybe change all events handling to pass through an event cache to handle temporarily disabled events. - evaluate the changes required for multi-process+shared mem or multi-thread +thread-local+fast locking. - ability to kill an arbitrary session from the command line. Put a "kill now" flag in every session which preempts any other processing and wake the session up. - ability to decide whether to drain or kill sessions when putting a server to maintenance mode => requires a per-server session list and the change above. Old, maybe obsolete points - clarify licence by adding a 'MODULE_LICENCE("GPL")' or something equivalent. - 3 memory models : failsafe (prealloc), normal (current), optimal (alloc on demand) - ability to assign a task priority based on L7 matching - implement support for event-triggerred epoll() - verify if it would be worth implementing an epoll_ctl_batch() for Linux - option minservers XXX : activates some spare servers when active servers are insufficient - new keyword 'check' : check http xxx, check smtp xxx, check ssl-hello