- Author(s): Mark D. Roth ([email protected])
- Approver: a11r
- Status: Implemented
- Implemented in: C-core (Go and Java in progress)
- Last updated: 2017-12-19
- Discussion at: https://groups.google.com/d/topic/grpc-io/DkweyrWEXxU/discussion
This document proposes a mechanism for encoding gRPC service config data in DNS for use in the open-source world.
The service config mechanism was originally designed for use inside of Google. However, all but one part of the original design works fine in the open-source world. That one part is the specification of how the service config data will be encoded in DNS. This proposal fills in this missing piece.
N/A
There are two parts to this proposal. The first part is to add some JSON wrapping for controlling how service config changes are canary tested. The second part describes how the service config is encoded in DNS.
When deploying a change to a service config, it is useful to be able to canary test changes to avoid wide-spread breakage by slowly increasing the number of clients that see the new version. To that end, multiple service configs choices can be listed, in order, along with criteria that determine which choice will be selected by a given client:
// A list of one or more service config choices.
// The first matching entry wins.
[
{
// Criteria used to select this choice.
// If a field is absent or empty, it matches all clients.
// All fields must match a client for this choice to be selected.
// If any unexpected field name is present in this object, the entire
// config is considered invalid.
//
// Client language(s): a list of strings (e.g., "c++", "java", "go",
// "python", etc). Each string is case insensitive.
"clientLanguage": [string],
// Percentage: integer from 0 to 100 indicating the percentage of
// clients that should use this choice. If present, the number must
// match the regular expression `^0|[0-9]|[1-9][0-9]|100$`
// All other numbers are considered invalid.
"percentage": number,
// Client hostname(s): a list of strings. Each name is case
// sensitive and must be an exact match of the hostname according to
// the system.
"clientHostname": [string],
// The service config data object for clients that match the above
// criteria. (The format for this object is defined in
// https://github.com/grpc/grpc/blob/master/doc/service_config.md.)
// If this field is not an object, or is missing, or is otherwise
// invalid, the entire config is considered invalid.
"serviceConfig": object
}
]
If the service config choice cannot be parsed, or otherwise is not semantically valid, the entire config MUST be discarded as per A21: Service Config Error Handling.
In DNS, the service config data (in the form documented in the previous
section) will be encoded in a TXT record via the mechanism described in
RFC-1464 using the attribute name
grpc_config
. The attribute value will be a JSON list containing service
config choices. The TXT record will be for a DNS name that is the same
as the gRPC server name, but with the prefix _grpc_config.
.
For example, here is an example TXT record for server myserver
:
_grpc_config.myserver 3600 TXT "grpc_config=[{\"serviceConfig\":{\"loadBalancingPolicy\":\"round_robin\",\"methodConfig\":[{\"name\":[{\"service\":\"MyService\",\"method\":\"Foo\"}],\"waitForReady\":true}]}}]"
Note that TXT records are limited to 255 bytes per string, as per RFC-1035 section 3.3. However, there can be multiple strings, which will be concatenated together, as described in RFC-4408 section 3.1.3. The total DNS response cannot exceed 65535 bytes. (See the "Open Issues" section below for more discussion.)
Note that because TXT records must be ASCII, this also imposes the restruction that the contents of the service config are also ASCII (e.g., service and method names, load balancing policy names, etc).
The service config is designed to be returned as part of name resolution, so encoding it in DNS makes the most sense. Sites that use a naming system other than DNS can, of course, implement their own resolvers with their own mechanism for encoding service config data.
When encoding the service config in DNS, TXT records are the "obvious" choice, since the service config is effectively additional metadata associated with the DNS name.
We use the _grpc_config
prefix for the DNS entry to allow a service
config to be specified for a service whose primary record is a CNAME
record, because DNS does not allow any other record for a same name
that contains a CNAME record.
The implementation has been done in C-core as part of the c-ares DNS resolver. We are currently working toward making the c-ares resolver the default DNS resolver for C-core. This requires things like Windows and Node support and adding address sorting.
DNS TXT records do have some limitations that need to be taken into account here. In particular:
- If a DNS response exceeds 512 bytes, it will fall back from UDP to TCP, which adds overhead.
- The total DNS response cannot exceed 65535 bytes.
- It is not clear whether individual DNS implementations will allow anywhere close to 65535 bytes, even though the spec says that they should.
Feedback is requested on whether these considerations will be a significant drawback for this design (in which case the design will probably have to be changed).