Copyright 2020 Google LLC Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.
Code generated by protoc-gen-go. DO NOT EDIT. versions: protoc-gen-go v1.25.0 protoc v3.13.0 source: google/monitoring/v3/common.proto
The tier of service for a Workspace. Please see the [service tiers documentation](https://cloud.google.com/monitoring/workspaces/tiers) for more details. Deprecated: Do not use.
An invalid sentinel value, used to indicate that a tier has not been provided explicitly.
The Stackdriver Basic tier, a free tier of service that provides basic features, a moderate allotment of logs, and access to built-in metrics. A number of features are not available in this tier. For more details, see [the service tiers documentation](https://cloud.google.com/monitoring/workspaces/tiers).
The Stackdriver Premium tier, a higher, more expensive tier of service that provides access to all Stackdriver features, lets you use Stackdriver with AWS accounts, and has a larger allotments for logs and metrics. For more details, see [the service tiers documentation](https://cloud.google.com/monitoring/workspaces/tiers).
The `Aligner` specifies the operation that will be applied to the data points in each alignment period in a time series. Except for `ALIGN_NONE`, which specifies that no operation be applied, each alignment operation replaces the set of data values in each alignment period with a single value: the result of applying the operation to the data values. An aligned time series has a single data value at the end of each `alignment_period`. An alignment operation can change the data type of the values, too. For example, if you apply a counting operation to boolean values, the data `value_type` in the original time series is `BOOLEAN`, but the `value_type` in the aligned result is `INT64`.
No alignment. Raw data is returned. Not valid if cross-series reduction is requested. The `value_type` of the result is the same as the `value_type` of the input.
Align and convert to [DELTA][google.api.MetricDescriptor.MetricKind.DELTA]. The output is `delta = y1 - y0`. This alignment is valid for [CUMULATIVE][google.api.MetricDescriptor.MetricKind.CUMULATIVE] and `DELTA` metrics. If the selected alignment period results in periods with no data, then the aligned value for such a period is created by interpolation. The `value_type` of the aligned result is the same as the `value_type` of the input.
Align and convert to a rate. The result is computed as `rate = (y1 - y0)/(t1 - t0)`, or "delta over time". Think of this aligner as providing the slope of the line that passes through the value at the start and at the end of the `alignment_period`. This aligner is valid for `CUMULATIVE` and `DELTA` metrics with numeric values. If the selected alignment period results in periods with no data, then the aligned value for such a period is created by interpolation. The output is a `GAUGE` metric with `value_type` `DOUBLE`. If, by "rate", you mean "percentage change", see the `ALIGN_PERCENT_CHANGE` aligner instead.
Align by interpolating between adjacent points around the alignment period boundary. This aligner is valid for `GAUGE` metrics with numeric values. The `value_type` of the aligned result is the same as the `value_type` of the input.
Align by moving the most recent data point before the end of the alignment period to the boundary at the end of the alignment period. This aligner is valid for `GAUGE` metrics. The `value_type` of the aligned result is the same as the `value_type` of the input.
Align the time series by returning the minimum value in each alignment period. This aligner is valid for `GAUGE` and `DELTA` metrics with numeric values. The `value_type` of the aligned result is the same as the `value_type` of the input.
Align the time series by returning the maximum value in each alignment period. This aligner is valid for `GAUGE` and `DELTA` metrics with numeric values. The `value_type` of the aligned result is the same as the `value_type` of the input.
Align the time series by returning the mean value in each alignment period. This aligner is valid for `GAUGE` and `DELTA` metrics with numeric values. The `value_type` of the aligned result is `DOUBLE`.
Align the time series by returning the number of values in each alignment period. This aligner is valid for `GAUGE` and `DELTA` metrics with numeric or Boolean values. The `value_type` of the aligned result is `INT64`.
Align the time series by returning the sum of the values in each alignment period. This aligner is valid for `GAUGE` and `DELTA` metrics with numeric and distribution values. The `value_type` of the aligned result is the same as the `value_type` of the input.
Align the time series by returning the standard deviation of the values in each alignment period. This aligner is valid for `GAUGE` and `DELTA` metrics with numeric values. The `value_type` of the output is `DOUBLE`.
Align the time series by returning the number of `True` values in each alignment period. This aligner is valid for `GAUGE` metrics with Boolean values. The `value_type` of the output is `INT64`.
Align the time series by returning the number of `False` values in each alignment period. This aligner is valid for `GAUGE` metrics with Boolean values. The `value_type` of the output is `INT64`.
Align the time series by returning the ratio of the number of `True` values to the total number of values in each alignment period. This aligner is valid for `GAUGE` metrics with Boolean values. The output value is in the range [0.0, 1.0] and has `value_type` `DOUBLE`.
Align the time series by using [percentile aggregation](https://en.wikipedia.org/wiki/Percentile). The resulting data point in each alignment period is the 99th percentile of all data points in the period. This aligner is valid for `GAUGE` and `DELTA` metrics with distribution values. The output is a `GAUGE` metric with `value_type` `DOUBLE`.
Align the time series by using [percentile aggregation](https://en.wikipedia.org/wiki/Percentile). The resulting data point in each alignment period is the 95th percentile of all data points in the period. This aligner is valid for `GAUGE` and `DELTA` metrics with distribution values. The output is a `GAUGE` metric with `value_type` `DOUBLE`.
Align the time series by using [percentile aggregation](https://en.wikipedia.org/wiki/Percentile). The resulting data point in each alignment period is the 50th percentile of all data points in the period. This aligner is valid for `GAUGE` and `DELTA` metrics with distribution values. The output is a `GAUGE` metric with `value_type` `DOUBLE`.
Align the time series by using [percentile aggregation](https://en.wikipedia.org/wiki/Percentile). The resulting data point in each alignment period is the 5th percentile of all data points in the period. This aligner is valid for `GAUGE` and `DELTA` metrics with distribution values. The output is a `GAUGE` metric with `value_type` `DOUBLE`.
Align and convert to a percentage change. This aligner is valid for `GAUGE` and `DELTA` metrics with numeric values. This alignment returns `((current - previous)/previous) * 100`, where the value of `previous` is determined based on the `alignment_period`. If the values of `current` and `previous` are both 0, then the returned value is 0. If only `previous` is 0, the returned value is infinity. A 10-minute moving mean is computed at each point of the alignment period prior to the above calculation to smooth the metric and prevent false positives from very short-lived spikes. The moving mean is only applicable for data whose values are `>= 0`. Any values `< 0` are treated as a missing datapoint, and are ignored. While `DELTA` metrics are accepted by this alignment, special care should be taken that the values for the metric will always be positive. The output is a `GAUGE` metric with `value_type` `DOUBLE`.
A Reducer operation describes how to aggregate data points from multiple time series into a single time series, where the value of each data point in the resulting series is a function of all the already aligned values in the input time series.
No cross-time series reduction. The output of the `Aligner` is returned.
Reduce by computing the mean value across time series for each alignment period. This reducer is valid for [DELTA][google.api.MetricDescriptor.MetricKind.DELTA] and [GAUGE][google.api.MetricDescriptor.MetricKind.GAUGE] metrics with numeric or distribution values. The `value_type` of the output is [DOUBLE][google.api.MetricDescriptor.ValueType.DOUBLE].
Reduce by computing the minimum value across time series for each alignment period. This reducer is valid for `DELTA` and `GAUGE` metrics with numeric values. The `value_type` of the output is the same as the `value_type` of the input.
Reduce by computing the maximum value across time series for each alignment period. This reducer is valid for `DELTA` and `GAUGE` metrics with numeric values. The `value_type` of the output is the same as the `value_type` of the input.
Reduce by computing the sum across time series for each alignment period. This reducer is valid for `DELTA` and `GAUGE` metrics with numeric and distribution values. The `value_type` of the output is the same as the `value_type` of the input.
Reduce by computing the standard deviation across time series for each alignment period. This reducer is valid for `DELTA` and `GAUGE` metrics with numeric or distribution values. The `value_type` of the output is `DOUBLE`.
Reduce by computing the number of data points across time series for each alignment period. This reducer is valid for `DELTA` and `GAUGE` metrics of numeric, Boolean, distribution, and string `value_type`. The `value_type` of the output is `INT64`.
Reduce by computing the number of `True`-valued data points across time series for each alignment period. This reducer is valid for `DELTA` and `GAUGE` metrics of Boolean `value_type`. The `value_type` of the output is `INT64`.
Reduce by computing the number of `False`-valued data points across time series for each alignment period. This reducer is valid for `DELTA` and `GAUGE` metrics of Boolean `value_type`. The `value_type` of the output is `INT64`.
Reduce by computing the ratio of the number of `True`-valued data points to the total number of data points for each alignment period. This reducer is valid for `DELTA` and `GAUGE` metrics of Boolean `value_type`. The output value is in the range [0.0, 1.0] and has `value_type` `DOUBLE`.
Reduce by computing the [99th percentile](https://en.wikipedia.org/wiki/Percentile) of data points across time series for each alignment period. This reducer is valid for `GAUGE` and `DELTA` metrics of numeric and distribution type. The value of the output is `DOUBLE`.
Reduce by computing the [95th percentile](https://en.wikipedia.org/wiki/Percentile) of data points across time series for each alignment period. This reducer is valid for `GAUGE` and `DELTA` metrics of numeric and distribution type. The value of the output is `DOUBLE`.
Reduce by computing the [50th percentile](https://en.wikipedia.org/wiki/Percentile) of data points across time series for each alignment period. This reducer is valid for `GAUGE` and `DELTA` metrics of numeric and distribution type. The value of the output is `DOUBLE`.
Reduce by computing the [5th percentile](https://en.wikipedia.org/wiki/Percentile) of data points across time series for each alignment period. This reducer is valid for `GAUGE` and `DELTA` metrics of numeric and distribution type. The value of the output is `DOUBLE`.
The typed value field. Types that are assignable to Value: *TypedValue_BoolValue *TypedValue_Int64Value *TypedValue_DoubleValue *TypedValue_StringValue *TypedValue_DistributionValue
A 64-bit double-precision floating-point number. Its magnitude is approximately ±10<sup>±300</sup> and it has 16 significant digits of precision.
A closed time interval. It extends from the start time to the end time, and includes both: `[startTime, endTime]`. Valid time intervals depend on the [`MetricKind`](https://cloud.google.com/monitoring/api/ref_v3/rest/v3/projects.metricDescriptors#MetricKind) of the metric value. In no case can the end time be earlier than the start time. * For a `GAUGE` metric, the `startTime` value is technically optional; if no value is specified, the start time defaults to the value of the end time, and the interval represents a single point in time. If both start and end times are specified, they must be identical. Such an interval is valid only for `GAUGE` metrics, which are point-in-time measurements. * For `DELTA` and `CUMULATIVE` metrics, the start time must be earlier than the end time. * In all cases, the start time of the next interval must be at least a millisecond after the end time of the previous interval. Because the interval is closed, if the start time of a new interval is the same as the end time of the previous interval, data written at the new start time could overwrite data written at the previous end time.
Optional. The beginning of the time interval. The default value for the start time is the end time. The start time must not be later than the end time.
Describes how to combine multiple time series to provide a different view of the data. Aggregation of time series is done in two steps. First, each time series in the set is _aligned_ to the same time interval boundaries, then the set of time series is optionally _reduced_ in number. Alignment consists of applying the `per_series_aligner` operation to each time series after its data has been divided into regular `alignment_period` time intervals. This process takes _all_ of the data points in an alignment period, applies a mathematical transformation such as averaging, minimum, maximum, delta, etc., and converts them into a single data point per period. Reduction is when the aligned and transformed time series can optionally be combined, reducing the number of time series through similar mathematical transformations. Reduction involves applying a `cross_series_reducer` to all the time series, optionally sorting the time series into subsets with `group_by_fields`, and applying the reducer to each subset. The raw time series data can contain a huge amount of information from multiple sources. Alignment and reduction transforms this mass of data into a more manageable and representative collection of data, for example "the 95% latency across the average of all tasks in a cluster". This representative data can be more easily graphed and comprehended, and the individual time series data is still available for later drilldown. For more details, see [Filtering and aggregation](https://cloud.google.com/monitoring/api/v3/aggregation).
The `alignment_period` specifies a time interval, in seconds, that is used to divide the data in all the [time series][google.monitoring.v3.TimeSeries] into consistent blocks of time. This will be done before the per-series aligner can be applied to the data. The value must be at least 60 seconds. If a per-series aligner other than `ALIGN_NONE` is specified, this field is required or an error is returned. If no per-series aligner is specified, or the aligner `ALIGN_NONE` is specified, then this field is ignored.
An `Aligner` describes how to bring the data points in a single time series into temporal alignment. Except for `ALIGN_NONE`, all alignments cause all the data points in an `alignment_period` to be mathematically grouped together, resulting in a single data point for each `alignment_period` with end timestamp at the end of the period. Not all alignment operations may be applied to all time series. The valid choices depend on the `metric_kind` and `value_type` of the original time series. Alignment can change the `metric_kind` or the `value_type` of the time series. Time series data must be aligned in order to perform cross-time series reduction. If `cross_series_reducer` is specified, then `per_series_aligner` must be specified and not equal to `ALIGN_NONE` and `alignment_period` must be specified; otherwise, an error is returned.
The reduction operation to be used to combine time series into a single time series, where the value of each data point in the resulting series is a function of all the already aligned values in the input time series. Not all reducer operations can be applied to all time series. The valid choices depend on the `metric_kind` and the `value_type` of the original time series. Reduction can yield a time series with a different `metric_kind` or `value_type` than the input time series. Time series data must first be aligned (see `per_series_aligner`) in order to perform cross-time series reduction. If `cross_series_reducer` is specified, then `per_series_aligner` must be specified, and must not be `ALIGN_NONE`. An `alignment_period` must also be specified; otherwise, an error is returned.
The set of fields to preserve when `cross_series_reducer` is specified. The `group_by_fields` determine how the time series are partitioned into subsets prior to applying the aggregation operation. Each subset contains time series that have the same value for each of the grouping fields. Each individual time series is a member of exactly one subset. The `cross_series_reducer` is applied to each subset of time series. It is not possible to reduce across different resource types, so this field implicitly contains `resource.type`. Fields not specified in `group_by_fields` are aggregated away. If `group_by_fields` is not specified and all the time series have the same resource type, then the time series are aggregated into a single output time series. If `cross_series_reducer` is not defined, this field is ignored.
The pages are generated with Goldsv0.3.2-preview. (GOOS=darwin GOARCH=amd64)
Golds is a Go 101 project developed by Tapir Liu.
PR and bug reports are welcome and can be submitted to the issue list.
Please follow @Go100and1 (reachable from the left QR code) to get the latest news of Golds.