davsclaus opened a new pull request, #29878:
URL: https://github.com/apache/camel/pull/29878

   ## What
   
   CAMEL-26938, a sub-task of CAMEL-26935 (make the Camel CLI faster for AI 
agents).
   
   A `camel cmd` action, a runtime tool of the MCP server, `camel ask` and the 
TUI talk to a running integration through files in `~/.camel`. Both sides 
polled every 100 millis, and the client slept 100 millis before it first 
looked, so an action took about 110 millis even when the integration answered 
at once.
   
   - `ActionClient` looks for the answer after 5 millis, and then waits a tenth 
of the time it has waited so far, from 5 to 100 millis. An answer is read at 
most about a tenth later than it is there, and an action that takes long, such 
as a message sent to a slow route, is not looked for 200 times a second.
   - The connector looks for action files every 20 millis (was 100). Listing 
`~/.camel` costs far more than looking at its modification time, so between two 
status updates it only lists the directory when that time has changed, which it 
does when a file is added to it. The status update every second lists it anyway.
   - A file added within the same tick of the file system clock as the change 
before it leaves the time as it was (Windows ticks about every 16 millis, and 
`File.lastModified` is in millis), so a time is only remembered once it is 20 
millis old: until then the next poll lists again.
   - An action file that cannot be read yet, such as one a Camel CLI before 
4.21 writes in place in the shared `<pid>-action.json`, makes the next poll 
list again, as finishing it in place leaves the directory as it is.
   - `ActionBaseCommand.getJsonObject` (the route structure and topology of 
source files) looks before it sleeps, as the dump is usually there already.
   - `ActionClient.executeText` takes an answer that is not JSON, of an 
integration of an earlier Camel version, once it has not changed for 100 
millis. It was: the same at two polls, which were 100 millis apart.
   - The unused `RuntimeHelper.POLL_INTERVAL_MS` is removed.
   
   The protocol is the same, so a new client is faster with an older 
integration, and the other way round.
   
   ## Measured
   
   macOS (M4 Pro), JDK 25. The client and the connector run in one warm JVM, 
with a home of their own, and send 100 actions with a pause of 37-86 millis 
between them (both runs shown):
   
   | client | connector | median | p90 |
   |---|---|---|---|
   | main | main | 112 ms | 114 ms |
   | this PR | this PR | 23 / 26 ms | 33 / 34 ms |
   | main | this PR | 111 ms | 114 ms |
   | this PR | main | 41 ms | 62 ms |
   
   CPU of an idle file transport over 15 seconds, the JVM included, with 67 
other files in `.camel`, on a machine with other load:
   
   | other integrations writing their status every second | main | this PR |
   |---|---|---|
   | 0 | 1.56 / 1.57 % | 1.55 / 1.24 % |
   | 5 | 1.53 / 1.46 % | 2.09 / 2.19 % |
   
   The status of another integration is moved in place every second, which 
changes the directory too, so with more integrations the poll lists more often. 
Listing at every poll, every 20 millis, would cost about 3 % of a core in a 
separate benchmark.
   
   ## Tests
   
   - `FileCliConnectorTransportTest`: between two status updates the directory 
is listed less than 15 times a second (the poll runs about 50 times), and a 
half-written action makes every poll list it again (10 times within a second). 
Each test fails without its change.
   - `ActionClientTest`: how long the client waits between two looks, and an 
answer that is not JSON is taken once it has not changed for 100 millis.
   - The tests of `camel-cli-connector` and `camel-jbang-core` pass.
   
   _Claude Code on behalf of Claus Ibsen (davsclaus)_
   
   🤖 Generated with [Claude Code](https://claude.com/claude-code)
   
   https://claude.ai/code/session_012nJsFsjqqFiwroeHRWaKbS
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to