HTTP
""""

Schéma de base
''''''''''''''

.. graphviz::

  digraph "arch" {
    nodesep=0.6;
    ranksep=1.7;
    subgraph cluster_client1 {
      label="Client 1";
      "http_client1"[label="http_client"];
    }
    subgraph cluster_client2 {
      label="Client 2";
      "http_client2"[label="http_client"];
    }
    subgraph cluster_proxy1 {
      label="Proxy 1"
      "haproxy_1"[label="haproxy"];
    }
    subgraph cluster_proxy2 {
      label="Proxy 2"
      "haproxy_2"[label="haproxy"];
    }
    subgraph cluster_httpd1 {
      label="Islet HTTP Server 1"
      "server1"[label="apache"];
    }
    subgraph cluster_httpd2 {
      label="Islet HTTP Server 2"
      "server2"[label="apache"];
    }
    subgraph cluster_httpd3 {
      label="Worker HTTP Server 1"
      "server3"[label="apache"];
    }
    subgraph cluster_httpd4 {
      label="Worker HTTP Server 2"
      "server4"[label="apache"];
    }
    subgraph cluster_httpd5 {
      label="Worker HTTP Server 3"
      "server5"[label="apache"];
    }
    "http_client1" -> "haproxy_1"[color=blue,weight=10]
    "http_client1" -> "haproxy_2"[color=blue,weight=10, style=dotted]
    "http_client2" -> "haproxy_2"[color=blue,weight=10]
    "http_client2" -> "haproxy_1"[color=blue,weigth=10, style=dotted]
    "http_client1" -> "haproxy_1"[color=blue,weight=10,dir=back]
    "http_client1" -> "haproxy_2"[color=blue,weight=10,style=dotted,dir=back]
    "http_client2" -> "haproxy_2"[color=blue,weight=10,dir=back]
    "http_client2" -> "haproxy_1"[color=blue,weigth=10,style=dotted,dir=back]

    "haproxy_1" -> server1 [color=green]
    "haproxy_1" -> server2 [color=green]
    "haproxy_1" -> server3 [color=green]
    "haproxy_1" -> server4 [color=green]
    "haproxy_1" -> server5 [color=green]

    "haproxy_2" -> server1 [color=green]
    "haproxy_2" -> server2 [color=green]
    "haproxy_2" -> server3 [color=green]
    "haproxy_2" -> server4 [color=green]
    "haproxy_2" -> server5 [color=green]

    "http_client1" -> server1 [color=red, dir=both]
    "http_client2" -> server2 [color=red, dir=both]

    "haproxy_1" -> "haproxy_2" [label=VRRP,constraint=false]

  { rank = sink;
    Legend [shape=none, margin=0, label=<
    <TABLE BORDER="0" CELLBORDER="1" CELLSPACING="0" CELLPADDING="4">
     <TR>
      <TD COLSPAN="2"><B>Legend</B></TD>
     </TR>
     <TR>
      <TD>Requête originale + redirect 302</TD>
      <TD BGCOLOR="BLUE"></TD>
     </TR>
     <TR>
      <TD>Checks</TD>
      <TD BGCOLOR="GREEN"></TD>
     </TR>
     <TR>
      <TD>Data</TD>
      <TD BGCOLOR="RED"></TD>
     </TR>
    </TABLE>
   >];
  }

  }

Description
'''''''''''

The HTTP service is of primary importance for the day-to-day operations of the computing system (updates, boot process, …). We therefore suggest to set up  the HTTP service in such a way that its only limiting factor is the network bandwith.

To do so, we combine `HAProxy`, `keepalived` and `apache`. Thus, the clients communicate with an unique URL pointing towards 2 PI addresses (DNS round robin). This set up offers a first level of load balancing.

The two IP addresses used are attributed to the 2 `HAProxy` machines. These are virtual IP addresses which are taken care of by a `keepalived` service, which is it running on both machines.  Please note that these machines should be located on different hypervisors. The proxy service will be provided by `HAProxy`, allowing for finer grained load balancing towards the final HTTPS servers.

For the http clients that support transparent http redirections, the `HAProxy` will not serve as proxy for the requests, but redirect the requests to the most appropriate server (identified by the load balancing algorithm). The other clients (`iPXE`+ `curl` mainly at this point in time)  will see the `HAProxy` as a transparent proxy. 

The final servers will run on the machines dedicated to the island management (with  higher network throughput) and on the `workers`. It will be the role of the load balancer to maximize the available bandwith onto the different available HTTP servers.

In oder to reduce the risk of cascading failures, we suggest to limit the number of requests sent to the workers. Thus, the service will be composed of: 

* N HTTPS serves, with N being the total number of hypervisors (TOP+WORKERS+2xISLET). Note that the data of the HTTP Server is stored in gluster file system, and not cached by apache. 
* 2 machines with `HAProxy` servers paired via `keepalived`.


There will be a single and unique URL for all clients. 
This service will host the package repository, the `bmgr` service and the services needed for deployment of the compute nodes (kickstart, diskless, …)



.. raw:: latex

  \clearpage