-
Notifications
You must be signed in to change notification settings - Fork 3
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
[SARC-329] Implémenter les alertes : Proportion de jobs CPU avec stat…
…s prometheus sur un noeud donné plus bas qu’un threshold X
- Loading branch information
1 parent
0e6ba5c
commit 82b843e
Showing
4 changed files
with
428 additions
and
7 deletions.
There are no files selected for viewing
177 changes: 177 additions & 0 deletions
177
sarc/alerts/usage_alerts/prometheus_stats_occurrences.py
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,177 @@ | ||
import logging | ||
from datetime import datetime, timedelta | ||
from typing import Dict, List, Optional, Sequence, Union | ||
|
||
from sarc.config import MTL | ||
from sarc.jobs.series import compute_time_frames, load_job_series | ||
|
||
logger = logging.getLogger(__name__) | ||
|
||
|
||
class PrometheusStatInfo: | ||
"""Prometheus stat context, used in checking below.""" | ||
|
||
def __init__(self, name): | ||
self.name = name | ||
self.col_has = f"has_{name}" | ||
self.col_ratio = f"ratio_{name}" | ||
self.avg = None | ||
self.stddev = None | ||
self.threshold = None | ||
|
||
|
||
def check_prometheus_stats_occurrences( | ||
time_interval: Optional[timedelta] = timedelta(days=7), | ||
time_unit=timedelta(days=1), | ||
minimum_runtime: Optional[timedelta] = timedelta(minutes=5), | ||
cluster_names: Optional[List[str]] = None, | ||
group_by_node: Optional[Sequence[str]] = ("mila",), | ||
min_jobs_per_group: Optional[Union[int, Dict[str, int]]] = None, | ||
nb_stddev=2, | ||
): | ||
""" | ||
Check if we have scrapped Prometheus stats for enough jobs per node per cluster per time unit. | ||
Log a warning for each node / cluster where ratio of jobs with Prometheus stats is lower than | ||
a threshold computed using mean and standard deviation statistics from all clusters. | ||
Parameters | ||
---------- | ||
time_interval: timedelta | ||
If given, only jobs which ran in [now - time_interval, time_interval] will be used for checking. | ||
Default is last 7 days. | ||
If None, all jobs are used. | ||
time_unit: timedelta | ||
Time unit in which we must check cluster usage through time_interval. Default is 1 day. | ||
minimum_runtime: timedelta | ||
If given, only jobs which ran at least for this minimum runtime will be used for checking. | ||
Default is 5 minutes. | ||
If None, set to 0. | ||
cluster_names: list | ||
Optional list of clusters to check. | ||
There may have clusters we don't want to check among retrieved jobs (eg. clusters in maintenance). | ||
On the opposite, we may expect to see jobs in a cluster, but there are actually no jobs in this cluster. | ||
To cover such cases, one can specify the complete list of expected clusters with `cluster_names`. | ||
Jobs from clusters not in this list will be ignored both to compute statistics and in checking phase. | ||
If a cluster in this list does not appear in jobs, a warning will be logged. | ||
If empty (or not specified), use all clusters available among jobs retrieved with time_interval. | ||
group_by_node: Sequence | ||
Optional sequence of clusters to group by node. | ||
For clusters in this list, we will check each node separately (ie. a "group" is a cluster node). | ||
By default, we check the entire cluster (i.e. the "group" is the cluster itself). | ||
min_jobs_per_group: int | dict | ||
Minimum number of jobs required for checking in each group. | ||
Either an integer, as minimum number for any group, | ||
or a dictionary mapping a cluster name to minimum number in each group of this cluster | ||
A group is either a cluster node, if cluster name is in `group_by_node`, | ||
or the entire cluster otherwise. | ||
Default is 1 job per group. | ||
nb_stddev: float | ||
Amount of standard deviation to remove from average statistics to compute checking threshold. | ||
Threshold is computed as: | ||
max(0, average - nb_stddev * stddev) | ||
""" | ||
|
||
# Parse time_interval and get data frame | ||
start, end, clip_time = None, None, False | ||
if time_interval is not None: | ||
end = datetime.now(tz=MTL) | ||
start = end - time_interval | ||
clip_time = True | ||
df = load_job_series(start=start, end=end, clip_time=clip_time) | ||
|
||
# Parse minimum_runtime, and select only jobs where | ||
# elapsed time >= minimum runtime and allocated.gres_gpu == 0 | ||
if minimum_runtime is None: | ||
minimum_runtime = timedelta(seconds=0) | ||
df = df[ | ||
(df["elapsed_time"] >= minimum_runtime.total_seconds()) | ||
& (df["allocated.gres_gpu"] == 0) | ||
] | ||
|
||
# List clusters | ||
cluster_names = cluster_names or sorted(df["cluster_name"].unique()) | ||
|
||
# Split data frame into time frames using `time_unit` | ||
df = compute_time_frames(df, frame_size=time_unit) | ||
|
||
# Duplicates lines per node to count each job for each node where it runs | ||
df = df.explode("nodes") | ||
|
||
# If cluster not in group_by_node, | ||
# then we must count jobs for the entire cluster, not per node. | ||
# To simplify the code, let's just define 1 common node for all cluster jobs | ||
cluster_node_name = "(all)" | ||
df.loc[~df["cluster_name"].isin(group_by_node), "nodes"] = cluster_node_name | ||
|
||
# Add a column to ease job count | ||
df.loc[:, "task_"] = 1 | ||
|
||
# Generate Prometheus context for each Prometheus stat we want to check. | ||
prom_contexts = [ | ||
PrometheusStatInfo(name=prom_col) | ||
for prom_col in ["cpu_utilization", "system_memory"] | ||
] | ||
|
||
# Add columns to check if job has prometheus stats | ||
for prom in prom_contexts: | ||
df.loc[:, prom.col_has] = ~df[prom.name].isnull() | ||
|
||
# Group per timestamp per cluster per node, and count jobs and prometheus stats. | ||
# If "cluster_names" are given, use only jobs in these clusters. | ||
f_stats = ( | ||
df[df["cluster_name"].isin(cluster_names)] | ||
.groupby(["timestamp", "cluster_name", "nodes"])[ | ||
[prom_info.col_has for prom_info in prom_contexts] + ["task_"] | ||
] | ||
.sum() | ||
) | ||
|
||
# Compute ratio of job with Prometheus stat for each group, | ||
# then compute threshold for each Prometheus stat. | ||
for prom in prom_contexts: | ||
f_stats[prom.col_ratio] = f_stats[prom.col_has] / f_stats["task_"] | ||
prom.avg = f_stats[prom.col_ratio].mean() | ||
prom.stddev = f_stats[prom.col_ratio].std() | ||
prom.threshold = max(0, prom.avg - nb_stddev * prom.stddev) | ||
|
||
# Parse min_jobs_per_group | ||
default_min_jobs = 1 | ||
if min_jobs_per_group is None: | ||
min_jobs_per_group = {} | ||
elif isinstance(min_jobs_per_group, int): | ||
default_min_jobs = min_jobs_per_group | ||
min_jobs_per_group = {} | ||
assert isinstance(min_jobs_per_group, dict) | ||
|
||
# Now we can check | ||
clusters_seen = set() | ||
for row in f_stats.itertuples(): | ||
timestamp, cluster_name, node = row.Index | ||
clusters_seen.add(cluster_name) | ||
nb_jobs = row.task_ | ||
if nb_jobs >= min_jobs_per_group.get(cluster_name, default_min_jobs): | ||
grouping_type = "cluster" if node == cluster_node_name else "node / cluster" | ||
grouping_name = ( | ||
f"[{cluster_name}]" | ||
if node == cluster_node_name | ||
else f"[{cluster_name}][{node}]" | ||
) | ||
for prom in prom_contexts: | ||
local_stat = getattr(row, prom.col_has) / nb_jobs | ||
if local_stat < prom.threshold: | ||
logger.warning( | ||
f"[{timestamp}]{grouping_name} insufficient Prometheus data for {prom.name}: " | ||
f"{round(local_stat * 100, 2)} % of CPU jobs / {grouping_type} / time unit; " | ||
f"minimum required: {prom.threshold} ({prom.avg} - {nb_stddev} * {prom.stddev}); " | ||
f"time unit: {time_unit}" | ||
) | ||
|
||
# Check clusters listed in `cluster_names` but not found in jobs. | ||
for cluster_name in cluster_names: | ||
if cluster_name not in clusters_seen: | ||
# No stats found for this cluster. Warning | ||
logger.warning( | ||
f"[{cluster_name}] no Prometheus data available: no job found" | ||
) |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Oops, something went wrong.