Skip to content

Latest commit

 

History

History
83 lines (72 loc) · 4.72 KB

DNS.md

File metadata and controls

83 lines (72 loc) · 4.72 KB
Raw

DNS in mtop

The mtop and mc binaries use DNS for both connecting to and discovering Memcached servers. They use both the system resolver and an internal DNS resolver based on the server hostnames provided. How that works and design decisions made for the internal DNS resolver are detailed below.

How DNS is used

If hostnames given to mtop or mc are prefixed with special strings, the internal DNS resolver is used to perform an initial resolution depending on the prefix. If the hostnames given to mtop are mc are not prefixed with the special strings described below, no initial resolution is performed.

If the hostname is prefixed with dns+, the hostname is resolved as A and AAAA records. The IP address of each record will be treated as a Memcached server. This resolution is only performed a single time at application start.

If the hostname is prefixed with dnssrv+, the hostname is resolved as a SRV record. The target field of each record will be treated as a Memcached server. The port field of the SRV record is ignored. Instead, the port portion of the server name from command line arguments is used. The target used for the Memcached server is resolved at connection time using the system resolver.

Internal DNS resolver

mtop and mc use an internal DNS resolver when given hostnames that are prefixed with special strings (dns+ and dnssrv+). How resolution is performed is described below.

  • Configuration is loaded from /etc/resolv.conf. Configuration is only loaded a single time at application start up. Only a subset of possible settings and options are supported. Notably, search domains are not supported. All supplied hostnames are assumed to be fully qualified. The supported settings and options are:
    • Top-level setting nameserver - An IP address of DNS server to use for resolution. If not set, 127.0.0.1 is used as a default.
    • Option timeout:n - Number of seconds after which a DNS request is considered timed out. If not set, 5 seconds is used as a default.
    • Option attempts:n - Total number of attempts to make to resolve a hostname. If not set, 2 attempts are made.
    • Option rotate - If set, use a different nameserver for each request instead of always starting with the first one. If not set, each request will always start with the first nameserver.
  • If the hostname has a special prefix, either A and AAAA resolutions of the hostname are performed or a SRV resolution is performed.
    • A dns+ prefix will cause A and AAAA resolutions to be performed. All results from these two DNS queries are used as Memcached servers with the port supplied as a command line argument. If using TLS, the hostname supplied on the command line is used to validate to the server certificate (unless overridden by a command line flag).
    • A dnssrv+ prefix will cause a SRV resolution to be performed. The target of each result is used as a Memcached server with the port supplied as a command line argument. If using TLS, the hostname supplied on the command line is used to validate the server certificate (unless override by a command line flag). If the target is a hostname, it will be resolved at connection time using the system resolver.
  • If the hostname does not have a special prefix, no resolution using the internal DNS resolver is done. Instead, the hostname or hostnames are each treated as individual servers. They will be resolved to A or AAAA records at connection time using the system resolver. If using TLS, the hostname supplied on the command line is used to validate the server certificate (unless overridden by a command line flag).

Limitations

These limitations apply when the internal DNS resolver is used.

  • resolv.conf is only loaded a single time at application start up. This usually won't be a problem in practice because the internal DNS resolver is only used for an initial lookup when using dns+ or dnssrv+ prefixes.
  • Search domains are not supported. Every domain is assumed to already be fully qualified.
  • EDNS is not supported. One implication of this is that only the default UDP packet size of 512 bytes is used. DNS responses that do not fit into a 512 byte UDP packet will cause TCP fallback to be used.
  • Results are not cached based on record TTLs or otherwise. This usually won't be a problem in practice because the internal DNS resolver is only used for an initial lookup when using dns+ or dnssrv+ prefixes.