"double free or corruption (!prev)" in XGBOOST4j

[cohenran]

In XGBOOST4j, I got this message after about 6 hours of running.
No other log was written.
I work on windows 10, on java 11

ml.dmlc xgboost4j_2.13 3.0.4

[trivialfis]

Seems like #11660 (comment) ? JVM + Windows. Do you have a JVM backtrace?

I don't have a guess at the moment.

[cohenran]

No, I don't have a backtrace.

[trivialfis]

ok, so what type of jobs are you running? Training, inference? Would you like to share some snippets that I can run?

[cohenran]

Right now I run training work.
I run the same training multiple times, to see which params are the best.
For instance, the params can be 0.05, 10, multi:softprob, logloss, 500 (eta, max depth, objective, eval metric, rounds).
I do either train or update:

	for (int i = 0; i < dataAndLabels.getSplittedLabels().length; i++) {
		DMatrix trainMat = createDMatrixSplit(dataAndLabels, i);

		Map<String, Object> paramsMap = setParams(params);

		if (i == 0) {
			booster = XGBoost.train(trainMat, paramsMap, params.getRounds(), new HashMap<>(), null, null);
		} else {
			booster.update(trainMat, dataAndLabels.getTrainingData().length);
		}
	}

public DMatrix createDMatrixSplit(DataAndLabels dataAndLabels, int i) throws XGBoostError, IOException {
	DMatrix trainMat = new DMatrix(
			dataAndLabels.getSplittedData()[i],
			dataAndLabels.getSplittedLabels()[i].length, 5, Float.NaN);
	trainMat.setLabel(dataAndLabels.getSplittedLabels()[i]);		

	return trainMat;
}

[trivialfis]

How did you install Java 11 and XGBoost? I suspect there's a mix of C runtimes here.

The message

double free or corruption (!prev)

Seems like the one from glibc instead Microsoft msvcrt or ucrt. Are you using a mingw toolchain?

[cohenran]

It was a java install on Ubuntu. I didn't install XGBoost - I simply added it in maven.
I don't know mingw toolchain. It is a simple java app.

java --version
openjdk 11.0.28 2025-07-15
OpenJDK Runtime Environment (build 11.0.28+6-post-Ubuntu-1ubuntu124.04.1)
OpenJDK 64-Bit Server VM (build 11.0.28+6-post-Ubuntu-1ubuntu124.04.1, mixed mode, sharing)

[trivialfis]

It was a java install on Ubuntu.

I thought you are using windows 10?

[cohenran]

Ohhh sorry, my mistake :(
The software runs on Ubuntu (The software runs on Ubuntu server. I use windows 10 locally, so thats was my mistake).
Sorry!

[trivialfis]

No worries. So, where did the error happen? Windows or Ubuntu? How about #11776 ?

[cohenran]

Both on the server Ubuntu.

[trivialfis]

I will try to recreate something that resembles #11775 (comment) and see if I can reproduce it next week.

In the meantime, if you could share a reproducible example, it would be of great help.

[cohenran]

Its hard, because it doesn't always happen.

[trivialfis]

At least some information like the shape of the input dataset would be useful. Without reproducing the issue there's no way I can guess the cause. So, anything that can help reproduce the issue is useful. We haven't seen segfault for a long time, aside from the issue I linked. So, I don't have any guess at the moment.

[cohenran]

6 ints as floats for every label.
Today all worked good.