Models can not be scaled up to 2 copies

[haiminh2001]

Context

I am learning how the auto scaling of the model mesh works. I found this piece of docs:

Models will scale to two copies if they have been used recently regardless of the load - the autoscaling behaviour applies between 2 and N>2 copies.

It is so vague that I have to dive in the source code, and then I found this code:

                                // For 1->2 copies, scale-up can also be triggered by a pattern of recent usage
                                // See explanation of CacheEntry#usageSlices
                                if (loadedCount == 1) {
                                    // assert mr.getInstanceIds().containsKey(instanceId);

                                    int i1 = ce.earlierUseIteration, i2 = ce.lastUsedIteration;
                                    // invariants: lower < upper, i1 <= i2
                                    if (logger.isDebugEnabled()) {
                                        logger.debug("Second copy trigger evaluation for model " + modelId
                                            + ": target range [" + lower + ", " + upper + "], I1="
                                                + i1 + ", I2=" + i2 + ", curIteration=" + iterationCounter);
                                    }
                                    boolean i1inRange = false, i2inRange = false;
                                    if (i2 >= lower && i1 <= upper) {
                                        i1inRange = i1 >= lower;
                                        i2inRange = i2 <= upper;
                                    }
                                    if (i2inRange || !i1inRange) {
                                        ce.earlierUseIteration = i2;
                                    }
                                    ce.lastUsedIteration = iterationCounter;

                                    if (i1inRange || i2inRange) {
                                        // Model was used within the target range [MIN_AGE, MAX_AGE] iterations ago
                                        // so trigger loading of a second copy

                                        // Don't do it if > 90% full and cache is younger than secondCopyLruThresholdMillis
                                        if ((10 * clusterStats.totalFree) / clusterStats.totalCapacity >= 1
                                                || (now - clusterStats.globalLru) > secondCopyLruThresholdMillis) {
                                            logger.info("Attempting to add second copy of model " + modelId
                                                    + " in another instance since \"regular\" usage was detected");
                                            ensureLoadedInternalAsync(modelId, lastTime, ce.getWeight(), excludeThisInstance, 0);
                                            continue;
                                        }
                                    }
                                }

As far as I understand, the logic if a model is recently used and there is a prior usage of it falling into the interval of 40 minutes and 7 minutes before the correspond time, the model should be scaled to 2 copies.

Current behaviour

If a model is consistently used, the earlierUseIteration and lastUsedIteration will be updated continuously, to the last check time. That logic is indicated in these lines of code:

                                    if (i2inRange || !i1inRange) {
                                        ce.earlierUseIteration = i2;
                                    }
                                    ce.lastUsedIteration = iterationCounter;

i2inRange and i1inRange will never be true, since both i1 i2 will always be updated to the most recently point of time and consequently exceed the upper point. Therefore a model has to be used once, wait for 7 minutes without receiving any requests in order to be scaled to 2 copies ( I have tested that behaviour). Having to wait for 7 minutes

Expectation

If a model is being used consistently for over 7 minutes, that model should be scaled to 2 copies.

Suggestion

Perhaps the point of time that the oldest request that not exceed 40 minutes should be recorded instead of the earlierUseIteration. The remaining logic is the same.