ResourceSpecification(mapping=None, *, ignore_unknown_fields=False, **kwargs)ResourceSpec collects a set of resources that can be used to specify requests and requirements.
Note: Highly experimental as this can be runtime dependent. Can use the "extras" field to experiment first before trying to abstract it.
Attributes |
|
|---|---|
| Name | Description |
cpu |
str
CPU specification. Examples: "100m", "0.5", "1", "2", ... correspond to 0.1, half, 1, or 2 cpus. Leave empty to let the system decide. Note that this does *not* determine the cpu vender/make, or its underlying clock speed and specific SIMD features. It is only the amount time it requires in timeslicing. |
cpu_limits |
str
CPU limit. Examples: "100m", "0.5", "1", "2", ... correspond to 0.1, half, 1, or 2 cpus. Leave empty to indicate no limit. |
memory |
str
Memory specification (in bytes). Examples: "128974848", "129e6", "129M", "123Mi", ... correspond to 128974848 bytes, 129000000 bytes, 129 mebibytes, 123 megabytes. Leave empty to let the system decide. |
memory_limits |
str
Memory usage limits. Examples: "128974848", "129e6", "129M", "123Mi", ... correspond to 128974848 bytes, 129000000 bytes, 129 mebibytes, 123 megabytes. Leave empty to indicate no limit. |
gpus |
int
Number of gpus. |
latency_budget_ms |
int
The maximum latency that this operator may use to process an element. If non positive, then a system default will be used. Operator developers should arrange for the system compute resources to be aligned with this latency budget; e.g. if you want a ML model to produce results within 500ms, then you should make sure you request enough cpu/gpu/memory to achieve that. |