CHANGES.md

Fast Changelog

Not yet released.

None yet.

v3.1.2

• Bump microtime to 3.0.0 for compatibility with node v6+

v3.1.1

• Added fsr_context.removeSocketEndListener() method so that long running RPC calls can remove their listener.

v3.1.0

• Added fsr_context.addSocketEndListener() method so that long running RPC calls can tell when the client socket has ended.

v3.0.1

• Add support for a metricLabels option when creating node-fast clients. This option must be a valid artedi labels object.

v3.0.0

• Bump the protocol version to 2 as a means to better deal with the CRC incompatibility stemming from the use of node-crc version 0.3.0. This change replaces the attempt at addressing this issue that was part of v2.8.1. That solution was found to be insufficient since there were cases where the calcluated CRC codes from the 0.3.0 version of node-crc were correct. With this release the protocol version is used to clearly determine which version of node-crc should be used to decode a received message and encode a response. Servers using this version are still compatible with fast clients using protocol version 1, but servers using protocol version 1 will not support clients using protocol version 2.

v2.8.1

• node-fast use wrong CRC library version when encoding errors to clients using updated CRC library version

v2.8.0

• node-fast provider transitive dependencies too old to build on x86_64

v2.7.0

• #23 Add support for updated crc package to improve interoperability

v2.6.1

• #22 Fix copy-paste bug in fast_server module

v2.6.0

• fast_client_request_time_ms became fast_client_request_time_seconds and is now measured in seconds rather than milliseconds.
• fast_request_time_ms became fast_server_request_time_seconds and is now measured in seconds rather than milliseconds.
• Add support for artedi v2 which requires histograms to use fixed buckets. When the collector passed is an artedi v2 collector, default buckets are generated via artedi.logLinearBuckets(10, -1, 3, 5). See: TritonDataCenter/node-artedi#17 for more details. Data generated with artedi v1 collectors are likely invalid (and have been with previous releases as well). Dashboards/queries for services which pass v2 collectors to node-fast, should restrict results using the {buckets_version="1"} label when performing Prometheus queries on this data.

v2.5.0

• #17 node-fast could track client metrics
• #16 fastserve cannot be run concurrently even with different arguments

v2.4.0

• #14 want node-fast FastServer 'onConnsDestroyed' method

v2.3.2

• #15 server leaks memory on socket errors
• #12 want support for node v6.12.0

v2.3.1

• #11 add RPC method name to metrics

v2.3.0

• #9 node-fast could track basic request metrics

v2.2.4

• #7 fsr_context isn't always unpiped from fsr_encoder, which causes memory leaks

v2.2.3

• #5 Server shutdown crashes when a connection had a socket error

v2.2.2

• #4 client should preserve "ase_errors" for Errors

v2.2.1

• #3 client should not log at "error" level

v2.2.0

• #2 update dependencies to support Node v4

v2.1.0

• #1 need limited support for servers that send null values

v2.0.0 (2016-08-29)

This is a complete rewrite of the Fast client and server implementation. Almost everything about the API has changed, including constructors, methods, and arguments. The primary difference is that the new Fast client is generally only responsible for Fast-protocol-level concerns. It does not do service discovery (via DNS), connection establishment, health monitoring, retries, or anything like that. Callers are expected to use something like node-cueball for that. The server manages a little bit more of the TCP state than the client does (particularly as clients connect), but callers are still expected to manage the server socket itself. To summarize: callers are expected to manage their own net.Socket (for clients) or net.Server (for servers), watching those objects for whatever events they're interested in and using those classes' methods for working with them. You no longer treat the Fast client as a wrapper for a net.Socket or the Fast server as a wrapper for a net.Server.

As a result of this change in design, the constructor arguments are pretty different. Many methods are gone (e.g., client.connect(), client.close(), server.listen(), server.address(), and so on). The interface is generally much simpler.

To make methods more extensible and consistent with Joyent's Best Practices for Error Handling (see README), the major APIs (constructors, client.rpc(), and server.rpc()) have been changed to accept named arguments in a single args argument.

Other concrete API differences include:

• Clients that make RPC calls get back a proper Readable stream, rather than an EventEmitter that emits message, end, and error events. (The previous implementation used the spirit of Node.js's object-mode streams, but predated their maturity, and so implemented something similar but ad-hoc.) As a result of this change, you cannot send null over the RPC channel, since object-mode streams use null to denote end-of-stream. Of course, you can wrap null in an object.
• Client.rpc() now takes named arguments as mentioned above.
• Server.rpc() is now called Server.registerRpcMethod() and takes named arguments as mentioned above.
• The server no longer provides an after event after RPC requests complete.

Other notable differences include:

• This implementation fixes a large number of protocol bugs, including cases where the protocol version was ignored in incoming messages and unsupported kinds of messages were treated as new RPC requests.
• This implementation does not support aborting RPC requests because the previous mechanism for this was somewhat dangerous. (See README.)
• The client DTrace probes have been revised. Server DTrace probes have been added. (See README.)
• Kang entry points have been added for observability. (See README.)

The original implementation can be found in the v0.x branch.