After long inactivity period, available token counter becomes negative

[MichelePaterniti]

I noticed that, at least with some bandwidth configuration, when a long time passes between subsequent calls to Bucket.tryConsume(...), the available token counter becomes negative and Bucket.tryConsume(...) return false, while there should be lots of tokens available.

I guess this could be caused by an arithmetic overflow in executing line 136
long divided = refillTokens * durationSinceLastRefillNanos + roundingError;
in private method io.github.bucket4j.BucketState.refill(int, Bandwidth, long, long).

To reproduce the problem, you can use the following test:

import java.time.Duration;

import org.junit.Test;

import io.github.bucket4j.Bandwidth;
import io.github.bucket4j.Bucket;
import io.github.bucket4j.Bucket4j;
import io.github.bucket4j.TimeMeter;

import static org.junit.Assert.assertTrue;

public class LongWaitTest {
    
    private final class TimeMeterWithArtificialDelay
            implements TimeMeter {
        private static final long serialVersionUID = 286427698074660314L;
        
        final long twelveHourNanos = 12 * 60 * 60 * 1_000_000_000L;
        private long delay = 0L;
        
        @Override
        public long currentTimeNanos() {
            return TimeMeter.SYSTEM_MILLISECONDS.currentTimeNanos() + delay;
        }
        
        void let12HoursPass() {
            delay = twelveHourNanos;
        }
    }
    
    @Test
    public void whenALongTimePassesThenAvailableTokensCounterBecomesNegative() throws InterruptedException {
        final Bandwidth limit1 = Bandwidth.simple(700000, Duration.ofHours(1));
        final Bandwidth limit2 = Bandwidth.simple(14500, Duration.ofMinutes(1));
        final Bandwidth limit3 = Bandwidth.simple(300, Duration.ofSeconds(1));
        final TimeMeterWithArtificialDelay customTimeMeter = new TimeMeterWithArtificialDelay();
        final Bucket bucket = Bucket4j.builder()
                .addLimit(limit1)
                .addLimit(limit2)
                .addLimit(limit3)
                .withCustomTimePrecision(customTimeMeter)
                .build();
        
        assertTrue(bucket.getAvailableTokens() > 0);
        assertTrue(bucket.tryConsume(1));
        
        customTimeMeter.let12HoursPass();
        
        assertTrue("Free tokens expected after long wait", bucket.getAvailableTokens() > 0);
        assertTrue("Free tokens expected after long wait", bucket.tryConsume(1));
    }
    
}

[vladimir-bukhtoyarov]

long divided = refillTokens * durationSinceLastRefillNanos + roundingError;

Yes, you properly detected the source of problem. I will release the fix in few days.

[vladimir-bukhtoyarov]

Which version do you use?

[MichelePaterniti]

I am using bucket4j-core 2.1.0 and bucket4j-jcache 2.1.0.

[vladimir-bukhtoyarov]

Looks like with your limits bucket will work normally only for 3,6 hours between invokation

public static void main(String[] args) {
        double d = Long.MAX_VALUE;
        System.out.println(d  / (700000L * 1_000_000_000L * 3600L));
    }

. As dirty work-around you can touch the bucket periodically:

if (bucket.tryConsume(1)) {
    bucket.addTokens(1);
}

The fixes will be prepared for branches 2.0, 2.1, 3.0, most likely tommorrow.

[vladimir-bukhtoyarov]

@MichelePaterniti please use 2.1.1(or 3.0.2) version.

[MichelePaterniti]

Thanks for the promptness of your fix.
It took us some time to integrate it and to run all tests and checks, but now it is in production and it is working fine.