SQL-like support on top of the CloudWatch SDK. Why that? Because...

1. It's fun to write parsers and query processors
2. Not a fan of the SDK for publishing and querying metrics
3. SQL is a language anyone is familar with
4. It's flexible in the sense that you can hook up anything you want on top of this (e.g., shell)

SELECT
  [nalias.]stat1 [as alias1],
  [nalias.]stat2 [as alias2],
  ...
FROM namespace [as nalias]
WHERE
  [nalias.]dimension1=value1,
  [nalias.]dimension2=value2
BETWEEN start_time AND end_time
PERIOD period

where statn is a statistic from a metric in the namespace given. Dimensions can be used for filtering also. The start and end times only support the ISO 8601 for now - e.g., 2018-08-01T02:10:38Z in the UTC time zone. The period is expressed in seconds. Here's an example of a full query:

SELECT
 max(CPUUtilization),
 avg(CPUUtilization)
FROM AWS/EC2
WHERE InstanceId='i-13k862a0e1h4673g6'
BETWEEN 2018-03-10T00:00:00Z
 AND 2018-03-10T14:00:00Z
PERIOD 3600

Select statements will translate to GetMetricStatistics requests to CloudWatch. One request per metric will be sent. So in the example above, even though two statistics are being selected, they refer to the same metric, so we can send a single request. A query like the above will return a result set with the max and avg joined by the timestamp, for example:

Note the column names are chosen based on the metric and statistics chosen. You can use aliases also - both for namespaces and the projections - to customize the final column names. So this query

SELECT
 max(CPUUtilization) as max_cpu,
 avg(CPUUtilization) as avg_cpu
FROM AWS/EC2 as ec2
WHERE ec2.InstanceId='i-13k862a0e1h4673g6'
BETWEEN 2018-03-10T00:00:00Z
 AND 2018-03-10T14:00:00Z
PERIOD 3600

would yield

The general format for publishing metrics is:

INSERT INTO namespace VALUES
  (metric_name_1, metric_unit_1, metric_value_1),
  ...
  (metric_name_n, metric_unit_n, metric_value_n),
 WITH
  dimension_name_1=dimension_value_1,
  ...
  dimension_name_n=dimension_value_n

For example, the statement below publishes 2 metrics (latency and buffer) tagged with 2 different dimensions each (InstanceId and InstanceType):

INSERT INTO Instances VALUES
  (latency, Milliseconds, 1234),
  (buffer, Kilobytes/Second, 98765)
 WITH
  InstanceID=123456,
  InstanceType=t2.micro

A query like that will translate to a single PutMetricData call to CloudWatch. The units have to be one of the units that CloudWatch recognizes, and query will fail at parsing time if that's not the case.