Isolates
/actuator/isolates shows the isolates of the application and how busy each is: the task manager as JSON. Spring Boot has no such endpoint.
Why
A slow server either computes or waits. The rows tell which: an isolate that computes runs late (lagMicros) and is busy (busy); one that waits for a database is neither, whatever its requests in flight. With a pool of server isolates the endpoint also shows whether the isolates share the work, and whether a task isolate is stuck.
Example
# lib/resources/application.properties
management.endpoints.web.exposure.include=health,isolates
$ curl localhost:8080/actuator/isolates
{"isolate": "main",
"isolates": [
{"name": "main", "kind": "main", "busy": 0.02, "lagMicros": 180,
"maxLagMicros": 2100, "active": 3, "memory": 96468992,
"details": {"requests": 18250}},
{"name": "isolate-1", "kind": "server", "busy": 0.03, "lagMicros": 210,
"maxLagMicros": 1900, "active": 2, "details": {"requests": 18102}}]}
What is listed
| Key | Content |
|---|---|
isolate |
The name of the isolate that answered this request |
isolates |
One row per isolate the application knows about |
The rows are those of the isolateLoads bean (see The load of the isolates):
- the application's first isolate,
main, with the memory of the process; - the isolates of the web server's pool,
isolate-1, …; - the isolates of the task executor,
task-1, …, while they exist; - what the application shows itself with
IsolateLoads.track, and the rows of its ownIsolateLoadSourcebeans.
| Field of a row | Meaning |
|---|---|
name |
The isolate's name, as in the logs |
kind |
main, server, task, engine, tracked, or a source's own word |
busy |
The share of the time it was working, from 0 to 1 |
lagMicros |
How late its event loop ran a timer, on average over the last seconds, in microseconds |
maxLagMicros |
The longest lag of that time |
active |
What it is working on now: requests in flight, tasks running |
memory |
The resident memory of the process in bytes; on the main row only, since isolates share it |
details |
What one source adds: requests (all the isolate has taken), refused (answered 503 because the server was full), tasks and queued of a task isolate |
A number a source cannot measure is left out.
Measuring starts with the first request
Nothing is measured while nobody asks. The first request starts the measuring, so its answer has no busy and no lag yet; ask again a few seconds later. Thirty seconds after the last request the measuring stops by itself.
So a monitor that asks every 10 seconds gets numbers from its second request on, and an application nobody watches pays nothing.
busy and the lag are estimates from inside the program, not processor time. What they can and cannot tell is on The load of the isolates.
With a pool of isolates
A request reaches any isolate, and so does a request for this endpoint. The answer is the same whichever isolate takes it:
- The endpoint asks the application's first isolate, which knows every isolate. The first isolate asks the others for their numbers and answers every row.
isolatenames the isolate that took the request, not the one that knows. It changes from one request to the next; the rows do not depend on it.- The rows of the other isolates are as fresh as that question: an isolate that does not answer within a second keeps its last row.
Exposure
The endpoint is enabled by default but, like every endpoint except health, not exposed over HTTP. Add isolates to management.endpoints.web.exposure.include, or use *:
management.endpoints.web.exposure.include=health,isolates
The rows show how loaded the server is, which is worth knowing to someone who wants to overload it: protect the endpoint (see Security). management.endpoint.isolates.access=none turns it off.
The endpoint is the bean isolatesEndpoint of spring_boot_starter_actuator; it backs off when the application declares an IsolatesEndpoint. It asks through the primaryIsolate bean (handler isolates.loads); see Asking the first isolate.
In Spring (Java)
Spring Boot Actuator has no counterpart: a JVM's threads share their beans. Nearest is the threaddump endpoint.