remove IgnoredTags since tagexclude achieves the same

table

handle text datatype and log unknown datatypes on table creation

Batches are created for every (table, columns) set as we may have metrics
with different set of tags & fields in the same batch.  This speeds up
inserts quite a bit by reducing the number of separate insert statements.

Grouping inserts requires columns to appear in the same order, so we now
sort them for both tags & fields.  This code could use a bit of tidying.

Make sure explicit schema is always used and allow changing that from
public into something else.

In some cases metrics datapoints only contain subset of all fields that
may be present for that metric. When storing columns separately instead
of using JSON getting new columns always resulted in the write failing
on first metric that was missing any columns and depending on how often
write was getting called and how many metrics had this behavior there
was potentially very long delay in getting all metrics through.

Used to cause SIGSEGV when the operation failed

The PostgreSQL plugin code is split up in multiple files for
better readability. Unit and Integration tests are written.
Code complexity reduced a bit by reducing branching.

A refactor was done that optimized several things:
1. Metrics are organized by measurement and schema validation is done
   only once instead of every metric
2. Metric values are batched and inserted with binary COPY
3. Tag IDs are now cached with a LRU strategy

commit 69fea1c broke tags_as_foreign keys when used with table_template. The template was being used for creation of the tag table when it should not be.

The previous code used a postgres serial for the tag table key. This results in telegraf having to lookup from the table to obtain the key. While there was an LRU cache to keep telegraf from having to constantly query, this posed a few problems.
1. The LRU cache is of a fixed size, but the performance can go from perfectly fine, to rock bottom as soon as you exceed the cache size. There is no gradual degredation that would allow the admin to detect the issue before it gets too bad. This has major repercussions as telegraf won't be able to keep up with ingest rate and will start dropping data.
2. The default for this cache was really small (1000), which is even likely too small for common use cases.
3. When telegraf starts up, the cache is cold, and so performance will be terrible while it is repopulated, again resulting in data loss.

The new solution solves this by using a deterministic tag key that does not require a lookup. The key is a 64-bit integer from the hash/maphash package. This hash is specifically designed for creating a hash of a string/bytes that can be used as a lookup key.

This is just a convenience so that when doing `select *`, you don't get a random order. This is not meant to be a perfect solution as any additional columns that are added to the table definition will not be sorted (this is just how postgres works). It's just meant so that the column ordering isn't a complete mess, making the life of the DBA somewhat better.

Previously the code used a mix of standard data types and extension types. This changes it to always use the standard postgres types. This addresses any inconsistencies that may arise due to the code not understanding that 2 types are really the same.

This switches the output to use a connection pool for managing the connection. Previously connection management was being done manually, accomplishing the same thing as if you had a pool size of 1. So this makes it much simpler.

This also removes the dbWrapper abstraction layer. There's no point to it when there's only one database driver. Plus having access to the underlying pool will make future work easier (without having to expand the abstraction, at which point it's no longer abstracting).

Previous code was querying the table structure prior to every insert. We should instead cache the table structure, and just assume the admin isn't going to do something silly like dropping columns or deleting the table while we're using it.

Previous code was performing a table structure lookup before every insert, which has an obvious performance impact.

However instead of just adding caching, the whole thing was ripped out and replaced. Bunch of reasons, but the new code is simpler, more memory efficient, and faster.

Previously the plugin only allowed very minor control over table templating. Could only adjust the table creation SQL for the main metrics table, and not the tag table. User also had no control over the ALTER TABLE statements.
This changes it so that all table CREATE & ALTER statements are configurable.

Additionally the templating was switched out from a rigid string substitution to Go templates. This allows a lot more flexibility in what can be done.

Lastly instead of only having a single statement per operation, requiring `;` to delimit multiple statements, the configuration now allows a list of statements, all run within a transaction. This is because pgx uses prepared statements, and postgres does not allow more than 1 command (command, not query) inside a prepared statement.

When schema updates are disabled, drop any metrics with tags that aren't in the database. We don't want to just throw an error as the rest of the metrics might be fine. We also don't want to insert without that tag as this can result in a collision with other metrics.

Had previously been removed while the code was being ripped apart. But now that the plugin is stabilizing, it's time to add it back. Still needs cleanup (note the TODO statements), but it's functional and moderately performant.

…ction mode

Some template methods filter columns based on role. For some columns, we can tell by the name (e.g. 'time' or 'tag_id'). However tag & field columns look the same. So we need a way to differentiate them. This change accomplishes that by identifying tag columns by using a PG column comment.

This is so that when users are looking up documentation on the templating functions, theyre not bombarded by lots of docs on internal methods that they don't care about.

Few other misc code cleanup things pulled in as well.

Since we send feedback to telegraf on error, it will retry the batch. Use a transaction to ensure that we don't insert the same data twice.

Remove the code that cleared caches on total DB connection loss. The intent was to be able to reset known state in the event of things like a database failover. But this is such an edge case, it's not worth the maintenance of the code.
If we ever want it back, it's here, in the git history.

This is a gray result. The point behind the go vet complaint is to prevent breakages when 3rd party packages update. However this isn't a third party package.

Also adds some slight documentation cleanup on the utils package.

These scenarios should never happen, but might as well just return an error instead.

golangci-lint is complaining that the package name shadows an import. This statement is incorrect, and no shaowing is happening. So renaming just to make the linter happy.

per-code-review, use testutil.MustMetric instead of metric.New

addresses linter warning

* origin/master: (183 commits)
  fix: CrateDB replace dots in tag keys with underscores (influxdata#9566)
  feat: Pull metrics from multiple AWS CloudWatch namespaces (influxdata#9386)
  fix: improve Clickhouse corner cases for empty recordset in aggregation queries, fix dictionaries behavior (influxdata#9401)
  fix(opcua): clean client on disconnect so that connect works cleanly (influxdata#9583)
  fix: Refactor ec2 init for config-api (influxdata#9576)
  fix: sort logs by timestamp before writing to Loki (influxdata#9571)
  fix: muting tests for udp_listener (influxdata#9578)
  fix: Do not return on disconnect to avoid breaking reconnect (influxdata#9524)
  fix: Fixing k8s nodes and pods parsing error (influxdata#9581)
  feat: OpenTelemetry output plugin (influxdata#9228)
  feat: Support AWS Web Identity Provider (influxdata#9411)
  fix: upgraded sensu/go to v2.9.0 (influxdata#9577)
  fix: Normalize unix socket path (influxdata#9554)
  docs: fix aws ec2 readme inconsistency (influxdata#9567)
  feat: Modbus Rtu over tcp enhancement (influxdata#9570)
  docs: information on new conventional commit format (influxdata#9573)
  docs: Add logo (influxdata#9574)
  docs: Adding links to net_irtt and dht_sensor external plugins (influxdata#9569)
  Upgrade hashicorp/consul/api to 1.9.1 (influxdata#9565)
  Update vmware/govmomi to v0.26.0 (influxdata#9552)
  Do not skip good quality nodes after a bad quality node is encountered (influxdata#9550)
  fix test so it hits a fake service (influxdata#9564)
  Update changelog
  Fix procstat plugin README to match sample config (influxdata#9553)
  Fix metrics reported as written but not actually written  (influxdata#9526)
  Prevent segfault in persistent volume claims (influxdata#9549)
  Update procstat to support cgroup globs & include systemd unit children (Copy of influxdata#7890) (influxdata#9488)
  Fix attempt to connect to an empty list of servers. (influxdata#9503)
  Fix handling bool in sql input plugin (influxdata#9540)
  Suricata alerts (influxdata#9322)
  Linter fixes for plugins/inputs/[fg]* (influxdata#9387)
  For Prometheus Input add ability to query Consul Service catalog (influxdata#5464)
  Support Landing page on Prometheus landing page (influxdata#8641)
  [Docs] Clarify tagging behavior (influxdata#9461)
  Change the timeout from all queries to per query (influxdata#9471)
  Attach the pod labels to the `kubernetes_pod_volume` & `kubernetes_pod_network` metrics. (influxdata#9438)
  feat(http_listener_v2): allows multiple paths and add path_tag (influxdata#9529)
  Bug Fix Snmp empty metric name (influxdata#9519)
  Worktable workfile stats (influxdata#8587)
  Update Go to v1.16.6 (influxdata#9542)
  ...