Add frequency tables to TSDB.

[tkaemming]

This adds a new data type, frequency tables, to the time series database and the API methods for:

• adding items,
• retrieving scores of known items over a time series (or aggregated over the series),
• retrieving the most frequent items seen over that a time series.

The notable implementation of this is implemented in the Redis backend, which uses Lua scripting to implement the frequency table data structure as a combination of a top-N index (modeled as a sorted set in Redis) and an estimation matrix for a Count-Min Sketch (implemented using a hash table). These two structures allows implementing estimated top-N queries of high (effectively unbounded) cardinality in almost fixed space. Counts are 100% accurate until the index is filled (and no extra space is used for the estimation matrix until this point), after which the data structure switches to a probabilistic implementation and accuracy begins to degrade for less frequently observed items, but remains accurate for more frequently observed items. Check out the comment in src/sentry/scripts/tsdb/cmsketch.lua for a more detailed explanation on implementation.

This adds 6 new metrics to the TSDB:

# number of events sent to server for an organization (key is always 0)
frequent_organization_received_by_system = 400
# number of events rejected by server for an organization (key is always 0)
frequent_organization_rejected_by_system = 401
# number of events blacklisted by server for an organization (key is always 0)
frequent_organization_blacklisted_by_system = 402
# number of events seen for a project, by organization
frequent_projects_by_organization = 403
# number of issues seen for a project, by project
frequent_issues_by_project = 404
# number of issues seen for a tag value, by issue:tag
frequent_values_by_issue_tag = 405

This depends on getsentry/rb#14, which adds a cluster.execute_commands method that can be used to run pipelines containing Lua scripts.

[tkaemming]

This is weird because there is an assumption in the API that all metrics have keys (in the case of the frequency table, a composite key) and I didn't want to overload the internal label with (0) another data type in case they end up having collisions.

[dcramer]

#2487

[codecov-io]

Current coverage is 82.72%

Merging #2637 into master will decrease coverage by -0.45% as of 8b04d56

@@            master   #2637   diff @@
======================================
  Files          862     864     +2
  Stmts        32699   32946   +247
  Branches         0       0       
  Methods          0               
======================================
+ Hit          27197   27256    +59
  Partial          0       0       
- Missed        5502    5690   +188

Review entire Coverage Diff as of 8b04d56

---

Uncovered Suggestions

1. +0.07% via ...try/utils/apidocs.py#414...436
2. +0.07% via ...try/utils/apidocs.py#117...139
3. +0.07% via ...gs/sentry_helpers.py#191...211
4. See 7 more...

Powered by Codecov. Updated on successful CI builds.