`service_name` label should determine series order

[kolesnikovae]

TL;DR: Let's rename service_name to __service_name__ because it is good for performance reasons.

---

Currently, we have a number of reserved labels (surrounded with __) that mainly determine the placement order of series in the block. The list is the following:

const (
	LabelNameProfileType = "__profile_type__"
	LabelNameType        = "__type__"
	LabelNameUnit        = "__unit__"
	LabelNamePeriodType  = "__period_type__"
	LabelNamePeriodUnit  = "__period_unit__"
	LabelNameDelta       = "__delta__"
	LabelNameProfileName = "__name__"
)

Cardinality of the label values is very low, a handful of unique combinations. As a result, series locality may be very poor, which leads to performance penalties due to the read amplification and redundant seek operations that we need to perform to fetch the data.

Consider an example: two programs app-1 and app-2 are being profiled , each of them is deployed to two pods. Then, their CPU profile series may form the following sequence (I omitted the reserved labels, because they are the same for each series and do not affect the order):

0: [a_label: bvc, pod_name: app-1-bvc, service_name: app-1]
1: [a_label: cxz, pod_name: app-2-cxz, service_name: app-2]
2: [a_label: vcx, pod_name: app-1-vcx, service_name: app-1]
3: [a_label: xcv, pod_name: app-2-xcv, service_name: app-2]

You can see that the series are interleaved: app-1, then app-2, app-1 again, and app-2 in the end. If we, say, want to query profiles of app-1 (rows 0 and 2), we either need to fetch more data (rows 0, 1, 2) or to skip the row 1 and seek to the row 2.

I assume that most of the queries cover a limited number of profiled programs (services/applications), usually just one. Therefore, from a performance standpoint, it would be beneficial to place profile series belonging to the same program close to each other. This would allow to read the data sequentially, and eliminate unnecessary "seeks", which is vital for queries from the object store where I/O latency reaches hundreds of milliseconds.

If we use __service_name__ label key instead of service_name, the order of series changes:

0: [__service_name__: app-1, a_label: bvc, pod_name: app-1-bvc]
1: [__service_name__: app-1, a_label: vcx, pod_name: app-1-vcx]
2: [__service_name__: app-2, a_label: cxz, pod_name: app-2-cxz]
3: [__service_name__: app-2, a_label: xcv, pod_name: app-2-xcv]

The order of the series is now perfect: profiles of the same service are stored sequentially in the block and can be fetched in a single pass.

[cyriltovena]

I'll take a stab at this.