[PATCH 6/9] vinyl: set range size automatically

[Vladimir Davydov]

On Mon, Jan 21, 2019 at 12:17:05AM +0300, Vladimir Davydov wrote:
> +int64_t
> +vy_lsm_range_size(struct vy_lsm *lsm)
> +{
> +	/* Use the configured range size if available. */
> +	if (lsm->opts.range_size > 0)
> +		return lsm->opts.range_size;
> +	/*
> +	 * It doesn't make much sense to create too small ranges.
> +	 * Limit the max number of ranges per index to 1000 and
> +	 * never create ranges smaller than 16 MB.
> +	 */
> +	enum { MIN_RANGE_SIZE = 16 * 1024 * 1024 };
> +	enum { MAX_RANGE_COUNT = 1000 };
> +	/*
> +	 * Ideally, we want to compact roughly the same amount of
> +	 * data after each dump so as to avoid IO bursts caused by
> +	 * simultaneous major compaction of a bunch of ranges,
> +	 * because such IO bursts can lead to a deviation of the
> +	 * LSM tree from the configured shape and, as a result,
> +	 * increased read amplification.  To achieve that, we need
> +	 * to have at least as many ranges as the number of dumps
> +	 * it takes to trigger major compaction in a range.
> +	 */
> +	int range_count = vy_lsm_dumps_per_compaction(lsm);

After having pondered this for a while, I'm inclined to think it isn't
such a good idea to use dumps_per_compaction for range_count: first, it
won't work for time series like workloads - the range_count won't scale
as the space size grows; second, after forcing compaction with
index.compact(), the dumps_per_compaction may collapse resulting in
massing range coalescing.

May be, we'd better use LSM tree fanout for the target number of ranges?
But how do we calculate it reliably?

> +	range_count = MIN(range_count, MAX_RANGE_COUNT);
> +	int64_t range_size = lsm->stat.disk.last_level_count.bytes /
> +						(range_count + 1);
> +	range_size = MAX(range_size, MIN_RANGE_SIZE);
> +	return range_size;
> +}