You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
We have identified a critical issue where there are significant timestamp inconsistencies between the recorded data for CPU, memory, and GPU metrics. Specifically, this issue arises when there are only a few data points available for CPU and memory metrics, while GPU metrics have a much higher frequency of recorded entries. For instance, in a recent workload, only two data points were available for CPU and memory over a given time range, whereas there were 30 entries for GPU. However, 28 of these GPU records did not have any corresponding CPU and memory data within a reasonable time proximity.
Impact:
When attempting to map these records, the absence of matching timestamps for CPU and memory results in the creation of new entries in the GPU map without any corresponding CPU and memory metrics. This leads to situations where the metrics map for a given pod contains only GPU values. Consequently, when performing pod-level calculations that rely on a complete set of metrics (CPU, memory, and GPU), the process encounters null values, causing errors or crashes.
The text was updated successfully, but these errors were encountered:
Description:
We have identified a critical issue where there are significant timestamp inconsistencies between the recorded data for CPU, memory, and GPU metrics. Specifically, this issue arises when there are only a few data points available for CPU and memory metrics, while GPU metrics have a much higher frequency of recorded entries. For instance, in a recent workload, only two data points were available for CPU and memory over a given time range, whereas there were 30 entries for GPU. However, 28 of these GPU records did not have any corresponding CPU and memory data within a reasonable time proximity.
Impact:
When attempting to map these records, the absence of matching timestamps for CPU and memory results in the creation of new entries in the GPU map without any corresponding CPU and memory metrics. This leads to situations where the metrics map for a given pod contains only GPU values. Consequently, when performing pod-level calculations that rely on a complete set of metrics (CPU, memory, and GPU), the process encounters null values, causing errors or crashes.
The text was updated successfully, but these errors were encountered: