lua-resty-upstream

Upstream connection load balancing and failover module

$ opm get hamishforbes/lua-resty-upstream

lua-resty-upstream

Upstream connection load balancing and failover module

Status

Experimental, API may change without warning.

Requires ngx_lua > 0.9.5

Overview

Create a lua shared dictionary. Define your upstream pools and hosts in init_by_lua, this will be saved into the shared dictionary.

Use the connect method to return a connected tcp socket.

Alternatively pass in a resty module (e.g lua-resty-redis or lua-resty-http) that implements connect() and set_timeout().

Call process_failed_hosts to handle failed hosts without blocking current request.

Use resty.upstream.api to modify upstream configuration during init or runtime, this is recommended!

resty.upstream.http wraps the lua-resty-http from @pintsized.

It allows for failover based on HTTP status codes as well as socket connection status.

    lua_shared_dict my_upstream_dict 1m;
    init_by_lua '
        upstream_socket  = require("resty.upstream.socket")
        upstream_api = require("resty.upstream.api")
    
        upstream, configured = upstream_socket:new("my_upstream_dict")
        if not upstream then
            error(configured)
        end
        api = upstream_api:new(upstream)
    
        if not configured then -- Only reconfigure on start, shared mem persists across a HUP
            api:create_pool({id = "primary", timeout = 100})
            api:set_priority("primary", 0)
            api:set_method("primary", "round_robin")
            api:add_host("primary", { id="a", host = "127.0.0.1", port = "80", weight = 10 })
            api:add_host("primary", { id="b", host = "127.0.0.1", port = "81",  weight = 10 })
    
            api:create_pool({id = "dr"})
            api:set_priority("dr", 10)
            api:add_host("dr", { host = "127.0.0.1", port = "82", weight = 5 })
            api:add_host("dr", { host = "127.0.0.1", port = "83", weight = 10 })
    
            api:create_pool({id = "test", priority = 5})
            api:add_host("primary", { id="c", host = "127.0.0.1", port = "82", weight = 10 })
            api:add_host("primary", { id="d", host = "127.0.0.1", port = "83", weight = 10 })
        end
    ';
    
    init_worker_by_lua 'upstream:init_background_thread()';
    
    server {
    
        location / {
            content_by_lua '
                local sock, err = upstream:connect()
                upstream:process_failed_hosts()
            ';
        }
    
    }

upstream.socket

new

syntax: upstream, configured = upstream_socket:new(dictionary, id?)

Returns a new upstream object using the provided dictionary name. When called in init_by_lua returns an additional variable if the dictionary already contains configuration. Takes an optional id parameter, this must be unique if multiple instances of upstream.socket are using the same dictionary.

init_background_thread

syntax: ok, err = upstream:init_background_thread()

Initialises the background thread, should be called in init_worker_by_lua

connect

syntax: ok, err = upstream:connect(client?, key?)

Attempts to connect to a host in the defined pools in priority order using the selected load balancing method. Returns a connected socket and a table containing the connected host, poolid and pool or nil and an error message.

When passed a socket or resty module it will return the same object after successful connection or nil.

Additionally, hash methods may take an optional key to define how to hash the connection to determine the host. By default ngx.var.remote_addr is used. This value is ignored when the pool's method is round robin.

    resty_redis = require('resty.redis')
    local redis = resty_redis.new()
    
    local key = ngx.req.get_headers()["X-Forwarded-For"]
    
    local redis, err = upstream:connect(redis, key)
    
    if not redis then
        ngx.log(ngx.ERR, err)
        ngx.status = 500
        return ngx.exit(ngx.status)
    end
    
    ngx.log(ngx.info, 'Connected to ' .. err.host.host .. ':' .. err.host.port)
    local ok, err = redis:get('key')

process_failed_hosts

syntax: ok, err = upstream:process_failed_hosts()

Processes any failed or recovered hosts from the current request. Spawns an immediate callback via ngx.timer.at, does not block current request.

get_pools

syntax: pools = usptream:get_pools()

Returns a table containing the current pool and host configuration. e.g.

    {
        primary = {
            up = true,
            method = 'round_robin',
            timeout = 100,
            priority = 0,
            hosts = {
                web01 = {
                    host = "127.0.0.1",
                    weight = 10,
                    port = "80",
                    lastfail = 0,
                    failcount = 0,
                    up = true,
                    healthcheck = true
                }
                web02 = {
                    host = "127.0.0.1",
                    weight = 10,
                    port = "80",
                    lastfail = 0,
                    failcount = 0,
                    up = true,
                    healthcheck = { interval = 30, path = '/check' }
                }
            }
        },
        secondary = {
            up = true,
            method = 'round_robin',
            timeout = 2000,
            priority = 10,
            hosts = {
                dr01 = {
                    host = "10.10.10.1",
                    weight = 10,
                    port = "80",
                    lastfail = 0,
                    failcount = 0,
                    up = true
                }
    
            }
        },
    }

save_pools

syntax: ok, err = upstream:save_pools(pools)

Saves a table of pools to the shared dictionary, pools must be in the same format as returned from get_pools

sort_pools

syntax: ok, err = upstream:sort_pools(pools)

Generates a priority order in the shared dictionary based on the table of pools provided

bind

syntax: ok, err = upstream:bind(event, func)

Bind a function to be called when events occur. func should expect 1 argument containing event data.

Returns true on a successful bind or nil and an error message on failure.

    local function host_down_handler(event)
        ngx.log(ngx.ERR, "Host: ", event.host.host, ":", event.host.port, " in pool '", event.pool.id,'" is down!')
    end
    local ok, err = upstream:bind('host_down', host_down_handler)

Event: host_up

Fired when a host changes status from down to up. Event data is a table containing the affected host and pool.

Event: host_down

Fired when a host changes status from up to down. Event data is a table containing the affected host and pool.

upstream.api

These functions allow you to dynamically reconfigure upstream pools and hosts

new

syntax: api, err = upstream_api:new(upstream)

Returns a new api object using the provided upstream object.

set_method

syntax: ok, err = api:set_method(poolid, method)

Sets the load balancing method for the specified pool. Currently randomised round robin and hashing methods are supported.

create_pool

syntax: ok, err = api:create_pool(pool)

Creates a new pool from a table of options, pool must contain at least 1 key id which must be unique within the current upstream object.

Other valid options are

  • method Balancing method

  • timeout Connection timeout in ms

  • priority Higher priority pools are used later

  • read_timeout

  • keepalive_timeout

  • keepalive_pool

  • status_codes See status_codes

Hosts cannot be defined at this point.

Note: IDs are converted to a string by this function

Default pool values

    { method = 'round_robin', timeout = 2000, priority = 0 }

set_priority

syntax: ok, err = api:set_priority(poolid, priority)

Priority must be a number, returns nil on error.

add_host

syntax: ok, err = api:add_host(poolid, host)

Takes a pool ID and a table of options, host must contain at least host. If the host ID is not specified it will be a numeric index based on the number of hosts in the pool.

Note: IDs are converted to a string by this function

Defaults:

    { host = '', port = 80, weight = 0}

remove_host

syntax: ok, err = api:remove_host(poolid, host)

Takes a poolid and a hostid to remove from the pool

down_host

syntax: ok,err = api:down_host(poolid, host)

Manually marks a host as down, this host will not be revived automatically.

up_host

syntax: ok,err = api:up_host(poolid, host)

Manually restores a dead host to the pool

upstream.http

Functions for making http requests to upstream hosts.

status_codes

This pool option is an array of status codes that indicate a failed request. Defaults to none.

The x character masks a digit

    {
        ['5xx'] = true, -- Matches 500, 503, 524
        ['400'] = true  -- Matches only 400
    }

new

syntax: httpc, err = upstream_http:new(upstream, ssl_opts?)

Returns a new http upstream object using the provided upstream object.

ssl_opts is an optional table for configuring SSL support.

  • ssl set to true to enable SSL Handshaking, default false

  • ssl_verify set to false to disable SSL certificate verification, default true

  • sni_host a string to use as the sni hostname, default is the request's Host header

    `lua https_upstream = Upstream_HTTP:new(upstream_ssl, { ssl = true, ssl_verify = true, sni_host = "foo.example.com" }) `

init_background_thread

syntax: ok, err = upstream_http:init_background_thread()

Initialises the background thread, should be called in init_worker_by_lua.

Do not call the init_background_thread method in upstream.socket if using the upstream.http background thread

request

syntax: res, err_or_conn_info, status? = upstream_api:request(params)

Takes the same parameters as lua-resty-http's request method.

On a successful request returns the lua-resty-http object and a table containing the connected host and pool.

If the request failed returns nil, the error and a suggested http status code

    local ok, err, status = upstream_http:request({
            path = "/helloworld",
            headers = {
                ["Host"] = "example.com",
            }
        })
    if not ok then
        ngx.status = status
        ngx.say(err)
        ngx.exit(status)
    else
        local host = err.host
        local pool = err.pool
    end

set_keepalive

syntax: ok, err = upstream_http:set_keepalive()

Passes the keepalive timeout / pool from the pool configuration through to the lua-resty-http set_keepalive method.

get_reused_times

syntax: ok, err = upstream_http:get_reused_times()

Passes through to the lua-resty-http get_reused_times method.

close

syntax: ok, err = upstream_http:close()

Passes through to the lua-resty-http close method.

HTTP Healthchecks

Active background healthchecks can be enabled by adding the healthcheck parameter to a host.

A value of true will enable the default check, a GET request for /.

The healthcheck parameter can also be a table of parameters valid for lua-resty-http's request method.

With a few additional parameters

  • interval to set the time between healthchecks, in seconds. Must be >= 10s. Defaults to 60s

  • timeout sets the connect timeout for healthchecks. Defaults to pool setting.

  • read_timeout sets the read timeout for healthchecks. Defaults to pool setting.

  • status_codes a table of invalid response status codes. Defaults to pool setting.

Failure for the background check is according to the same parameters as for a frontend request, unless overriden explicitly.

    -- Custom check parameters
    api:add_host("primary", {
         host = 123.123.123.123,
         port = 80,
         healthcheck = {
            interval = 30, -- check every 30s
            timeout      = (5*1000), -- 5s connect timeout
            read_timeout = (15*1000), -- 15s connect timeout
            status_codes = {["5xx"] = true, ["403"] = true}, -- 5xx and 403 responses are a fail
            -- resty-http params
            path = "/check",
            headers = {
                ["Host"] = "domain.com",
                ["Accept-Encoding"] = "gzip"
            }
         }
    })
    
    -- Default check parameters
    api:add_host("primary", {host = 123.123.123.123, port = 80, healthcheck = true})
    

TODO

  • IP based sticky sessions

  • Slow start - recovered hosts have lower weighting

  • Active TCP healthchecks

  • Use Cap'n Proto instead of JSON for serialisation

  • HTTP Minimum Rises - Hosts must have n succesful healthchecks before being marked up

  • HTTP Specific options

    • Cookie based sticky sessions

Authors

Hamish Forbes

License

mit

Dependencies

Versions