mkm85 / m-digest Goto Github PK

WIP, NOT PRODUCTION READY

The m-digest is a fast streaming quantile calculation algorithm especially targeted for resource constrained devices. The name and algorithm is inspired by t-digest/q-digest. [https://github.com/tdunning/t-digest].

• It uses a static array of centroids, this avoids malloc.
• All the algotihmic aspects which has to do with removing and inserting centroids into a set of centroids.

• Implement averaging between centroids when calculating quantiles.

The datastructure consists of m buckets where m is some fixed constant. Each bucket is a centroid. A centroid consists of a mean and a count.

The initial datastructure has a capacity of one for each bucket.

Every now and then the algorithm increases the capacity of the buckets. The capacity increase is made in such a way that the observations with the most relevance is in the smallest buckets.

Take the mdigest with most elements and merge the other mdigest into this, one bucket at a time. Start with the smallest bucket. At most a few increases is needed since we are basing the merge on the largest of the mdigests.

Given a mdigest, after increase_max_count is called the the fillrate is >= 1/6 implying that the fill rate was >= 1/3 before increase_max_count as it at most doubles the capacity.

lets assume that if two buckets can be merged to a lower bucket they are merged into this bucket. Such that the structure is optimally filled given the merge down algorithm.

bucket[i].count + bucket[i+1].count > bucket[i].maxCount, else they would be merged together into bucket[i].

if we assume bucket[i+1].maxCount = bucket[i].maxCount * 2 we get that bucket[i].maxCount + bucket[i+1].maxCount = 3*bucket[i].maxCount

this implies that the fillrate for bucket[i] and bucket[i+1] > bucket[i].maxCount / 3*bucket[i].maxCount = 1/3

this implies that the fill rate for a structure is atleast 1/3. We are not going to call increase_max_buckets before the structure is full, hence the structure is atleast 1/6 filled after a call to increase_max_count.

After a call to increase_max_count we still have the invariant that the fillrate for bucket i an i+1 is > 1/3 except for the last two buckets.

Since the last buckets are the largest, new insert search from the back until the best bucket for a fit is found. This way a binary search is not needed and the average number og searched buckets is log_2(number of buckets). which is as fast as a binary search for the desired bucket. This algorithm has a worst case complexity of O(m * n)

lets say we have two mdigests one for 24hr and one for 24hr and 5 minutes we want to know the quantiles for the 5 minute period. But since the 5 minute quantile data could all be around the 24hr median value and about zero data is store regarding the actual quantiles for these 5 minutes.

Instead collect several mdigests and merge them together for larger timespans.