SLO/HAProxy
Status: draft
Organizational
Service
This service is the HAProxy daemons used as TLS terminator as part of WMF’s production CDN edge infrastructure. This only covers the main haproxy daemon (/usr/sbin/haproxy) in its capacity to accurately and functionally serve production HTTP requests in real time; it does not cover the various ancillary tools, binaries, statistics/logging mechanisms, etc. All production instances in all datacenters are covered, which means all of the hardware machines currently named cp[0-9]*.{site}.wmnet.
Teams
SRE/Traffic owns and operates this service layer, and additionally the same team is also responsible for all of the direct dependencies at both edges of this service: the L4LB in the outwards-facing direction, and the Varnish frontend caches in the inwards-facing direction. Therefore, while this service impacts many other teams and services, the responsibility for it is fairly clearly a single-team affair. Other SRE subteams additionally share the burden of on-call incident response for this service.
Architectural
Environmental dependencies
This service runs independently in all 6 datacenters, and also comprises two different clusters named text and upload. In any given DC, text and upload have identical hardware configurations and layouts. We can (should?) define SLOs both for the global aggregate of a cluster (which may also need to discount manually-depooled time windows for specific datacenters, as with the discussion above about L4LB depools within a DC?), and per-DC per-cluster. The physical characteristics differ per-DC as follows (these numbers are for a single cluster text or upload):
| Datacenter | DC Layout | Cluster Machines | Cluster Layout |
|---|---|---|---|
| eqiad | 4 rows, multiple racks each | 8 | 2 machines per row, each in a distinct rack |
| codfw | 4 rows, multiple racks each | 8 | 2 machines per row, each in a distinct rack |
| esams | 1 row, 3 racks | 8 | Text - 3:3:2 Upload - 2:3:3 (machines in each of the 3 racks) |
| ulsfo | 1 row, 2 racks | 8 | 4:4 |
| eqsin | 1 row, 2 racks | 8 | 4:4 |
| drmrs | 1 row, 2 racks | 8 | 4:4 |
Service dependencies
Hard dependencies: Varnish Frontend caches. HAProxy is only able to return error messages if varnish isn't reachable.
Client-facing
Clients
Strictly speaking the L4LB layer would be the sole client of HAProxy. But considering L7 clients, every human and bot trying to reach Wikipedia or any other service hosted behind WMF's CDN acts a client of HAProxy
Service Level Indicators (SLIs)
Proxy combined latency-availability SLI: The percentage of all requests where the latency contributed by the proxy, excluding backend wait time, is within 50 milliseconds (excluding connection establishment and TLS handshake) and the request receives a non-error response or an error response originating at the backend. Errors originating at the proxy are measured as sessions with a termination state set to:
- I (Internal Error)
- R (Resource exhausted)
- D (Session killed by HAProxy)
- U (Session killed by HAProxy acting as a backup server)
- K (Session actively killed by an admin operating on HAProxy
- L (Session was locally processed by HAProxy and was not passed to a server)
See more details about termination states
Operational
Monitoring
The service is made of a supervisor haproxy process responsible for starting a child. The latter handles actual traffic, the former handles administrative commands and restarts the child if it stops responding. There are Icinga checks in place verifying that both RSA and ECDSA certificates kind be used to establish a TLS connection and that OCSP stapling material is up-to-date.
Troubleshooting
In most cases of anomalous operation it is sufficient to restart the service and open a task for the Traffic team with a description of the symptoms.
Deployment
The service is deployed by Puppet using the haproxy puppet module.
Service Level Objectives
The agreed-upon SLO is 99.9% of requests spending less than 50 milliseconds processing time inside of HAProxy itself and without a HAProxy Internal error.