influxdb/query/influxql

InfluxQL Transpiler

The InfluxQL Transpiler exists to rewrite an InfluxQL query into its equivalent query in Flux. The transpiler works off of a few simple rules that match with the equivalent method of constructing queries in InfluxDB.

NOTE: The transpiler code is not finished and may not necessarily reflect what is in this document. When they conflict, this document is considered to be the correct way to do it. If you wish to change how the transpiler works, modify this file first.

1. Identify the cursors
2. Identify the query type
3. Group the cursors
4. Create the cursors for each group
   1. Identify the variables
   2. Generate each cursor
   3. Join the cursors
   4. Evaluate the condition
   5. Perform the grouping
   6. Evaluate the function
   7. Combine windows
5. Join the groups
6. Map and eval columns
7. Encoding the results

Identify the cursors

The InfluxQL query engine works by filling in variables and evaluating the query for the values in each row. The first step of transforming a query is identifying the cursors so we can figure out how to fill them correctly. A cursor is any point in the query that has a variable or a function call. Math functions do not count as function calls and are handled in the eval phase.

For the following query, it is easy to identify the cursors:

SELECT max(usage_user), usage_system FROM telegraf..cpu

max(usage_user) and usage_system are the cursors that we need to fill in for each row. Cursors are global and are not per-field.

Identify the query type

There are three types of queries: raw, aggregate, and selector. A raw query is one where all of the cursors reference a variable. An aggregate is one where all of the cursors reference a function call. A selector is one where there is exactly one function call that is a selector (such as max() or min()) and the remaining variables, if there are any, are variables. If there is only one function call with no variables and that function is a selector, then the function type is a selector.

Group the cursors

We group the cursors based on the query type. For raw queries and selectors, all of the cursors are put into the same group. For aggregates, each function call is put into a separate group so they can be joined at the end.

Create the cursors for each group

We create the cursors within each group. This process is repeated for every group.

Identify the variables

Each of the variables in the group are identified. This involves inspecting the condition to collect the common variables in the expression while also retrieving the variables for each expression within the group. For a function call, this retrieves the variable used as a function argument rather than the function itself.

If a wildcard is identified, then the schema must be consulted for all of the fields and tags. If there is a wildcard in the dimensions (the group by clause), then the dimensions are excluded from the field expansion. If there is a specific listing of dimensions in the grouping, then those specific tags are excluded.

Generate each cursor

The base cursor for each variable is generated using the following template:

create_cursor = (db, rp="autogen", start, stop=now(), m, f) => from(bucket: db+"/"+rp)
    |> range(start: start, stop: stop)
    |> filter(fn: (r) => r._measurement == m and r._field == f)

Join the cursors

After creating the base cursors, each of them is joined into a single stream using an inner_join.

TODO(jsternberg): Raw queries need to evaluate fill() at this stage while selectors and aggregates should not.

> SELECT usage_user, usage_system FROM telegraf..cpu WHERE time >= now() - 5m
val1 = create_cursor(db: "telegraf", start: -5m, m: "cpu", f: "usage_user")
val1 = create_cursor(db: "telegraf", start: -5m, m: "cpu", f: "usage_system")
inner_join(tables: {val1: val1, val2: val2}, except: ["_field"], fn: (tables) => {val1: tables.val1, val2: tables.val2})

If there is only one cursor, then nothing needs to be done.

Evaluate the condition

At this point, generate the filter call to evaluate the condition. If there is no condition outside of the time selector, then this step is skipped.

Perform the grouping

We group together the streams based on the GROUP BY clause. As an example:

> SELECT mean(usage_user) FROM telegraf..cpu WHERE time >= now() - 5m GROUP BY time(5m), host
... |> group(by: ["_measurement", "host"]) |> window(every: 5m, ignoreGlobalBounds: true)

If the GROUP BY time(...) doesn't exist, window() is skipped. If there is no GROUP BY clause, it always groups by _measurement. If a wildcard is used for grouping, then this step is skipped. We also add ignoreGlobalBounds to every invocation of window() so the boundaries aren't clamped by the range() call.

Evaluate the function

If this group contains a function call, the function is evaluated at this stage and invoked on the specific column. As an example:

> SELECT max(usage_user), usage_system FROM telegraf..cpu
val1 = create_cursor(db: "telegraf", start: -5m, m: "cpu", f: "usage_user")
val1 = create_cursor(db: "telegraf", start: -5m, m: "cpu", f: "usage_system")
inner_join(tables: {val1: val1, val2: val2}, except: ["_field"], fn: (tables) => {val1: tables.val1, val2: tables.val2})
    |> max(column: "val1")

For an aggregate, the following is used instead:

> SELECT mean(usage_user) FROM telegraf..cpu
create_cursor(db: "telegraf", start: -5m, m: "cpu", f: "usage_user")
    |> group(except: ["_field"])
    |> mean(timeSrc: "_start", columns: ["_value"])

If the aggregate is combined with conditions, the column name of _value is replaced with whatever the generated column name is.

Combine windows

If there a window operation was added, we then combine each of the function results from the windows back into a single table.

... |> window(every: inf, ignoreGlobalBounds: true)

This step is skipped if there was no window function.

Join the groups

If there is only one group, this does not need to be done and can be skipped.

If there are multiple groups, as is the case when there are multiple function calls, then we perform an outer_join using the time and any remaining partition keys.

Map and eval the columns

After joining the results if a join was required, then a map call is used to both evaluate the math functions and name the columns. The time is also passed through the map() function so it is available for the encoder.

result |> map(fn: (r) => {_time: r._time, max: r.val1, usage_system: r.val2})

This is the final result. It will also include any tags in the partition key and the time will be located in the _time variable.

TODO(jsternberg): The _time variable is only needed for selectors and raw queries. We can actually drop this variable for aggregate queries and use the _start time from the partition key. Consider whether or not we should do this and if it is worth it.

Encoding the results

Each statement will be terminated by a yield() call. This call will embed the statement id as the result name. The result name is always of type string, but the transpiler will encode an integer in this field so it can be parsed by the encoder. For example:

result |> yield(name: "0")

The edge nodes from the query specification will be used to encode the results back to the user in the JSON format used in 1.x. The JSON format from 1.x is below:

{
    "results": [
        {
            "statement_id": 0,
            "series": [
                {
                    "name": "_measurement",
                    "tags": {
                        "key": "value"
                    },
                    "columns": [
                        "time",
                        "value"
                    ],
                    "values": [
                        [
                            "2015-01-29T21:55:43.702900257Z",
                            2
                        ]
                    ]
                }
            ]
        }
    ]
}

The measurement name is retrieved from the _measurement column in the results. For the tags, the values in the partition key that are of type string are included with both the keys and the values mapped to each other. Any values in the partition key that are not strings, like the start and stop times, are ignored and discarded. If the _field key is still present in the partition key, it is also discarded. For all normal fields, they are included in the array of values for each row. The _time field will be renamed to time (or whatever the time alias is set to by the query).

The chunking options that existed in 1.x are not supported by the encoder and should not be used. To minimize the amount of breaking code, using a chunking option will be ignored and the encoder will operate as normal, but it will include a message in the result so that a user can be informed that an invalid query option was used. The 1.x format has a field for sending back informational messages in it already.

TODO(jsternberg): Find a way for a column to be both used as a tag and a field. This is not currently possible because the encoder can't tell the difference between the two.