This PR replaces all char * in RRD structures with STRING *.

Once this is done in all structures and is merged, it is expected to have a big impact on the memory footprint on busy servers, since STRING * deduplicate the strings. So, all these contexts, dimensions, variables, etc that have the same name across dimensions, charts, nodes, alarms, etc will use the memory only once.

• rrdfamily
• rrddim
• rrdset
• rrdcalc
• alarm_entry
• alert_config
• rrdcalctemplate
• rrdvar
• rrddimvar
• rrdsetvar
• eval_variable
• rrdhost
• workers_utilization

Most of the changes are trivial (mostly a search/replace).

Also, this PR replaces AVL trees with DICTIONARIES for the following:

• rrddim
• rrdset
• rrdhost
• rrdvar
• rrdfamily

This change was required, because with the STRINGs we now don't need hashes to compare them, but in order to lookup a value in the AVL tree, we had to create a STRING, which resulted in thousands of string lookup per second on the string dictionary, which made things slower. By replacing AVL with DICTIONARIES, we can now enjoy all the benefits of string deduplication, without any performance penalties.

The following indexes were not useful and have been removed:

• rrdhost alarms_idx_name
  This was only needed during loading from disk on startup. It was replaced with a temporary dictionary.
• rrdhost alarms_health_log_idx
  This was only needed to find if an alarm entry of an rrdcalc is repeating. It is now turned to a flag of the alarm entries for the index is not needed.

Although linked lists should be replaced by dictionaries, which also support reference counting, this still needs some work. Still all the code base relies on linked lists.

Until we eliminate them, most of the linked lists are now double linked, to provide constant insertion and deletion times:

• rrdhost
• rrdset
• rrddim
• rrddimvar
• rrdsetvar
• rrdcalc
• rrdcalctemplate
• workers utilization

For the double linked list management, there are 5 very useful macros in libnedata.h, so this is an attempt to standardize them a bit.

The value for users is important:

With the current implementation, on a single node, strings are allocated 10 times on the average (6k entries, referenced 60k times). I expect the final to be around 20 times. The impact on busy parent agents will be really great.

We also saved about 200 bytes per chart, because now the RRDSET.id is not preallocated.

On busy parents we observed several GBs in memory reduction of the agent.

Other fixes

1. Chart labels are now written to SQLite a lot faster (one query per chart, instead of one query per label).
2. Stream compression function restored - we shouldn't have now disconnects.
3. On streaming start, the child first sends all chart and dimension definitions, before it starts streaming metric data.
4. Workers utilization now supports custom metric types, either ABSOLUTE or INCREMENTAL. Already used to monitor streaming receiver bandwidth and streaming sender bandwidth and circular buffer utilization ratio.
5. Fixed bug when charts were OBSOLETEd before any data collections and immediately were freed by the service thread.
6. RRDCONTEXT initialization are now using SQLite prepared statements to be faster.
7. RRDCONTEXT garbage collection was found not operational. Now it runs also on every agent restart and it has been rewritten to actually cleanup in one run.
8. RRDCONTEXT triggers are now monitored (there is a chart for this in the netdata section)
9. STRINGs are now monitored (there is a chart for this in the netdata section)
10. STRINGs are now lockless. Only the Judy hashtable is protected with a read/write lock - all operations on reference counters are using atomic operations to work lock free.
11. SQLite3 queries are now monitored (there is a chart for this in the netdata section)
12. streaming receivers were attempting to execute received commands without reason. Now they first check if there are any outstanding data for it.
13. RRDCONTEXT was infinitely re-triggering instance and context updates when a chart without retention on a parent, was streamed from a child (BEGIN, END) without setting values. Fixed: avoided the streaming of the empty chart, rrdcontext avoided the trigger on empty instances.
14. RRDCONTEXT hooks about collected metrics and instances have been moved after RRDDIMs and RRDSETs have data written to the database, so that RRDCONTEXT can find their retention.