ACDc Your friendly automatic call distribution commander

First, the "commander" part is a stretch, I know.

High level concepts involved here are queues and agents. There's a many-many relationship between queues and agents: a queue can have many agents, and an agent can be part of many queues. As such, there needs to be a distinct differentiation between what an agent is responsible for and what a queue tracks.


An agent represents an account user with a device (and possibly many devices). The agent tracks:

  1. what state it is in (and whether it can bridge member calls to the device)
  2. when it last had a call (some queue types are interested)
  3. the endpoint(s) to bridge a member's call to
  4. the current call (if agent is engaged by a member)
  5. the agent id
  6. the agent's account db

Agent Statuses

**init**: the agent process is starting up. Will not accept member connection requests.

**sync**: Waits for agent status event to know what state to transition to out of **sync**.

**ready**: the agent process is awaiting member connection requests.

**ringing**: the agent process has received a member connection win and is currently dialing the agent's device.

**answered**: the agent device has successfully bridged to the member. Waiting for the call to hangup.

**wrapup**: configurable wait time after a call has ended before the agent is ready to accept another call.

**paused**: the agent is on break or otherwise not taking calls. An optional timer can be set to limit how long the agent may remain in the paused state.
Agent Process Initialization

An agent can be represented by one or more Erlang processes, on any number of nodes, and in any one of the statuses above. As such, startup of the agent process is a tricky matter.

An agent process that has started up cannot just jump into the ready status, as the agent could be on the phone, away on break, or otherwise in-disposed. So before the process can leave the init status, the agent process published a sync_req message for all agent processes monitoring the agent. Any existing agent processes will receive the sync_req and reply with their current status. The intializing agent process will collect responses (or none) for a specified timeout and either sync with the "most active" agent process or, if no responses are received, enter the ready status.

If the sync_resp is in the init status as well, the agent process will move to ready to avoid deadlocking the two (or more) agent processes.

If the sync_resp is in the ready status, the initializing agent will enter the ready status as well. Member requests will now come to both (or all) agent processes.

If the sync_resp is in the waiting status, the initializing agent process will wait the wait_timeout time (how long a process waits to see if it 'won' the member), and send another sync_req.

If a sync_resp is in the ringing status, the initializing agent process will wait the wait_timeout time and send the sync_req again.

If a sync_resp is in the answered status, the response will include the call_id of the call being vied for (and possibly the b-leg call_id as well). The initializing agent will enter the same state and monitor the call for hangup. Upon receiving the hangup, the initializing agent process will enter the wrapup status, and after the wrap-up timeout, will enter the ready status (as other agent processes should too).

If a sync_resp is in the wrapup status, the resp should include the time left (otherwise use the full amount). The initializing agent will enter the wrapup status with the calculated timeout, then transition to ready.

If a sync_resp is in the paused status, the resp should include the time left (if on a timed break). If no time is included, the initializing agent process will enter the paused status indefinitely, until an agent_status event instructs the agent process to come back to ready.

Happy case for a member call

agent status: ready Agent process receives a member_connect_req message off the message bus. Agent publishes a member_connect_resp to the requesting queue's AMQP queue and enters the waiting state.

agent status: waiting Agent process receives member_connect_win, including the kapps_call json payload. Agent process publishes a bridge_req to the call control process. Agent publishes agent_state_change to other agent processes monitoring this agent_id. Agent process enters the ringing state.

agent status: ringing Agent process receives (eventually) the successful bridge event. Agent enters the answered state. Send the queue a member_connected message.

agent status: answered Agent process monitors the call's event stream, waiting for a hangup event. Upon reception, enter the wrapup state.

agent status: wrapup Queue may optionally specify a wrapup_timeout, delaying how long an agent must wait before returning to ready.

Q = A queue in an account AX(Y)(S) = Agent with id X, process Y, status S Status: (R)eady, (A)nswered, (W)aiting, (Ri)nging, (Wr)apup, (P)aused

An attempt at a message flowchart

1. Q sends a member_connect_req to all agents bound for the queue
                            ----> A1(1)(R)
Q ---> member_connect_req --|---> A1(2)(R)
                            ----> A2(1)(R)

2. Q recvs member_connect_resps from the two A1 processes and the A2 process
                             ---- A1(1)(W)
Q <--- member_connect_resp <-|--- A1(2)(W)
                             ---- A2(1)(W)

3a. Q decides A1(1) wins the member_connect and sends the member_connect_win
    A1(1) tries to ring the agent endpoint
Q ---> member_connect_win ----> A1(1)(Ri)

3b. Q sends A1(2) a member_connect_monitor to let it know an agent process has won the member_connect
    A1(2) enters the **ringing** status without dialing the endpoint
Q ---> member_connect_monitor ----> A1(2)(Ri)

3c. Q sends the A2(1) process a member_connect_ignore so it knows it failed to win the member.
    A2(1) re-enters the **ready** status
Q ---> member_connect_ignore ----> A2(2)(R)

4a. A1(\_) receives a failed dial to agent endpoint, alerts Q
    A1(\_) notes failed dial; if threshold reached, agent goes to **paused**, else **wrapup**
Q <--- member_retry --- A1(_)(Wr | P)

4b. A1(\_) receives bridged event, alerts Q
    A1(\_) enters the **answered** status
Q <--- member_connected --- A1(_)(A)

5. A1(\_) receives hangup, alerts Q
   A1(\_) enters the **wrapup** status
Q <--- member_hungup --- A1(_)(Wr)

5. Agent's wrapup timer expires, enters **ready** status
Q                        A1(_)(R)


Queues manage the dispersal of member calls to agents. Because the queue won't know if an agent is busy in another queue, the queue will broadcast a member_connect_req to all known agents. The queue will collect member_connect_resps and choose one agent, based on routing strategy, to send the member_connect_win message. If the agent is unable to connect the call and send the queue a member_connected, the agent will attempt the next appropriate member_connect_resp. If the list is exhausted, the queue will wait a configurable amount of time before sending another member_connect_req and repeating the process.

Queue Process Initialization

As with agents, multiple queue processes can (and should) be started per acdc queue. The challenge becomes how to coordinate member calls being distributed to agents with multiple queue processes running for the same queue.

Unlike agents, we don't want to consume as quickly as possible from the AMQP queue. We want a 'blocking' read from the AMQP queue while we process the caller. As soon as the caller is connected to the agent, the queue process should ACK the member_call message. Because of this constraint, we need a shared AMQP queue with auto_ack disabled.

Startup of a queue process should declare the AMQP queue as acdc.queue.ACCT.QID and set the auto_ack flag to false (will require manually ACK or NACK of the AMQP payload). Becoming a consumer of the queue will put the queue process into round-robin with other queue processes. At that point, the queue process is ready to receive the next member_call.

Of course, communication with the agent processes will occur with the queue processes dedicated AMQP queue. So each queue process will consume from both the shared AMQP queue and from their own private AMQP queue.

Queue Call Handling

1. member_call published from callflow module to shared AMQP queue acdc.queue.ACCT.QID
CF -- member_call --> Q1(1)

2a. Q1(1) recvs the member_call request, publishes the member_connect_req to the agents
Q1(1) -- member_connect_req --> As

3. Q1(1) recvs the member_connect_resps from the agents and decides (based on agent routing strategy such as round-robin or most-idle) which agent resp to declare victorious.
Q1(1) <-- member_connect_resp -- A

4a. Q1(1) sends the member_connect_win to the "winning" agent
Q1(1) -- member_connect_win --> A

4b. Q1(1) sends the member_connect_monitor to other agent processes that match the winning agent id
Q1(1) -- member_connect_monitor --> A

4c. Q1(1) sends the member_connect_ignore to all other responders
Q1(1) -- member_connect_ignore --> As

5. Q1(1) recvs the member_connected from the agent process. Q1(1) ACKs the member_call payload (and the next AMQP payload from #1 is delivered).
Q1(1) <-- member_connected -- A
  \------ ACK --------------> AMQP Shared Queue

5. Q1(\*) recvs member_hungup from agent (makes sure any stats triggered by the hangup are collected).
Q1(\*) <-- member_hungup -- A

6. If the agent process sends the member_connect_retry, it is published and consumed like the member_call message (meaning goto 2 and substitute member_call with member_connect_retry.

Edit this page here