def test_aggregation():
# note: uneven aggregation: we only want 1 resulting metric,
query = Query("")
query["avg_by"] = {"server": [""]}
query["sum_by"] = {"type": [""]}
targets = {
"web1.db": {"id": "web1.db", "tags": {"server": "web1", "type": "db", "n3": "foo"}},
"web1.php": {"id": "web1.php", "tags": {"server": "web1", "type": "php", "n3": "foo"}},
"web2.db": {"id": "web2.db", "tags": {"server": "web2", "type": "db", "n3": "foo"}},
"web2.php": {"id": "web2.php", "tags": {"server": "web2", "type": "php", "n3": "foo"}},
"web2.memcache": {"id": "web2.memcache", "tags": {"server": "web2", "type": "memcache", "n3": "foo"}},
}
from pprint import pprint
for (k, v) in targets.items():
v = Target(v)
v.get_graph_info(group_by={})
targets[k] = v
graphs, _query = build_graphs_from_targets(targets, query)
# TODO: there should be only 1 graph, containing all 5 items
print "Graphs:"
for (k, v) in graphs.items():
print "graph key"
pprint(k)
print "val:"
pprint(v)
assert {} == graphs
def build_graphs_from_targets(targets, query):
graphs = {}
if not targets:
return (graphs, query)
group_by = query['group_by']
sum_by = query['sum_by']
avg_by = query['avg_by']
avg_over = query['avg_over']
# i'm gonna assume you never use second and your datapoints are stored with
# minutely resolution. later on we can use config options for this (or
# better: somehow query graphite about it)
# note, the day/week/month numbers are not technically accurate, but
# since we're doing movingAvg that's ok
averaging = {
'M': 1,
'h': 60,
'd': 60 * 24,
'w': 60 * 24 * 7,
'mo': 60 * 24 * 30
}
if avg_over is not None:
avg_over_amount = avg_over[0]
avg_over_unit = avg_over[1]
if avg_over_unit in averaging.keys():
multiplier = averaging[avg_over_unit]
query['target_modifiers'].append(
Query.graphite_function_applier('movingAverage', avg_over_amount * multiplier))
# for each group_by bucket, make 1 graph.
# so for each graph, we have:
# the "constants": tags in the group_by
# the "variables": tags not in the group_by, which can have arbitrary
# values, or different values from a group_by tag that match the same
# bucket pattern
# go through all targets and group them into graphs:
for _target_id, target_data in sorted(targets.items()):
# FWIW. has an 'id' which timeserieswidget doesn't care about
target = Target(target_data)
target['target'] = target['id']
(graph_key, constants) = target.get_graph_info(group_by)
if graph_key not in graphs:
graph = {'from': query['from'], 'until': query['to']}
graph.update({'constants': constants, 'targets': []})
graphs[graph_key] = graph
graphs[graph_key]['targets'].append(target)
# ok so now we have a graphs dictionary with a graph for every appropriate
# combination of group_by tags, and each graph contains all targets that
# should be shown on it. but the user may have asked to aggregate certain
# targets together, by summing and/or averaging across different values of
# (a) certain tag(s). let's process the aggregations now.
if (sum_by or avg_by):
for (graph_key, graph_config) in graphs.items():
graph_config['targets_sum_candidates'] = {}
graph_config['targets_avg_candidates'] = {}
graph_config['normal_targets'] = []
for target in graph_config['targets']:
sum_id = target.get_agg_key(sum_by)
if sum_id:
if sum_id not in graph_config['targets_sum_candidates']:
graphs[graph_key]['targets_sum_candidates'][sum_id] = []
graph_config['targets_sum_candidates'][sum_id].append(target)
for (sum_id, targets) in graph_config['targets_sum_candidates'].items():
if len(targets) > 1:
for t in targets:
graph_config['targets'].remove(t)
graph_config['targets'].append(
graphite_func_aggregate(targets, sum_by, "sumSeries"))
for target in graph_config['targets']:
# Now that any summing is done, we look at aggregating by
# averaging because avg(foo+bar+baz) is more efficient
# than avg(foo)+avg(bar)+avg(baz)
# aggregate targets (whether those are sums or regular ones)
avg_id = target.get_agg_key(avg_by)
if avg_id:
if avg_id not in graph_config['targets_avg_candidates']:
graph_config['targets_avg_candidates'][avg_id] = []
graph_config['targets_avg_candidates'][avg_id].append(target)
for (avg_id, targets) in graph_config['targets_avg_candidates'].items():
if len(targets) > 1:
for t in targets:
graph_config['targets'].remove(t)
graph_config['targets'].append(
graphite_func_aggregate(targets, avg_by, "averageSeries"))
# remove targets/graphs over the limit
graphs = graphs_limit_targets(graphs, query['limit_targets'])
# Apply target modifiers (like movingAverage, summarize, ...)
for (graph_key, graph_config) in graphs.items():
for target in graph_config['targets']:
for target_modifier in query['target_modifiers']:
target_modifier(target, graph_config)
# if in a graph all targets have a tag with the same value, they are
# effectively constants, so promote them. this makes the display of the
# graphs less rendundant and makes it easier to do config/preferences
# on a per-graph basis.
for (graph_key, graph_config) in graphs.items():
# get all variable tags throughout all targets in this graph
tags_seen = set()
for target in graph_config['targets']:
for tag_name in target['variables'].keys():
tags_seen.add(tag_name)
# find effective constants from those variables,
# and effective variables. (unset tag is a value too)
first_values_seen = {}
effective_variables = set() # tags for which we've seen >1 values
for target in graph_config['targets']:
for tag_name in tags_seen:
# already known that we can't promote, continue
if tag_name in effective_variables:
continue
tag_value = target['variables'].get(tag_name, None)
if tag_name not in first_values_seen:
first_values_seen[tag_name] = tag_value
elif tag_value != first_values_seen[tag_name]:
effective_variables.add(tag_name)
effective_constants = tags_seen - effective_variables
# promote the effective_constants by adjusting graph and targets:
graph_config['promoted_constants'] = {}
for tag_name in effective_constants:
graph_config['promoted_constants'][tag_name] = first_values_seen[tag_name]
for target in graph_config['targets']:
target['variables'].pop(tag_name, None)
# now that graph config is "rich", merge in settings from preferences
constants = dict(graph_config['constants'].items() + graph_config['promoted_constants'].items())
for graph_option in get_action_on_rules_match(preferences.graph_options, constants):
if isinstance(graph_option, dict):
graph_config.update(graph_option)
else:
graph_config = graphs[graph_key] = graph_option(graph_config)
# but, the query may override some preferences:
override = {}
if query['statement'] == 'lines':
override['state'] = 'lines'
if query['statement'] == 'stack':
override['state'] = 'stacked'
if query['min'] is not None:
override['yaxis'] = override.get('yaxis', {})
override['yaxis'].update({'min': convert.parse_str(query['min'])})
if query['max'] is not None:
override['yaxis'] = override.get('yaxis', {})
override['yaxis'].update({'max': convert.parse_str(query['max'])})
graphs[graph_key].update(override)
# now that some constants are promoted, we can give the graph more
# unique keys based on all (original + promoted) constants. this is in
# line with the meaning of the graph ("all targets with those constant
# tags"), but more importantly: this fixes cases where some graphs
# would otherwise have the same key, even though they have a different
# set of constants, this can manifest itself on dashboard pages where
# graphs for different queries are shown.
# note that we can't just compile constants + promoted_constants,
# part of the original graph key is also set by the group by (which, by
# means of the bucket patterns doesn't always translate into constants),
# we solve this by just including the old key.
new_graphs = {}
for (graph_key, graph_config) in graphs.items():
new_key = ','.join('%s=%s' % i for i in graph_config['promoted_constants'].items())
new_key = '%s__%s' % (graph_key, new_key)
new_graphs[new_key] = graph_config
graphs = new_graphs
return (graphs, query)
def test_nontrivial_implicit_aggregation():
# we ultimately want 1 graph with 1 line for each server,
# irrespective of the values of the other tags (n1 and n2)
# and even whether or not the metrics have those tags at all.
query = Query("")
query['group_by'] = {}
query['sum_by'] = {'n1': [''], 'n2': ['']}
targets = {
# web1 : one with and without n2
'web1.a.a': {
'id': 'web1.a.a',
'tags': {
'server': 'web1',
'n1': 'a',
'n2': 'a'
}
},
'web1.a': {
'id': 'web1.a',
'tags': {
'server': 'web1',
'n1': 'a',
}
},
# web 2: 2 different values of n2
'web2.a.a': {
'id': 'web2.a.a',
'tags': {
'server': 'web2',
'n1': 'a',
'n2': 'a'
}
},
'web2.a.b': {
'id': 'web2.a.b',
'tags': {
'server': 'web2',
'n1': 'a',
'n2': 'b'
}
},
# web3: with and without n2, diff value for n1
'web3.a.a': {
'id': 'web3.a.a',
'tags': {
'server': 'web3',
'n1': 'a',
'n2': 'a'
}
},
'web3.b': {
'id': 'web3.b',
'tags': {
'server': 'web3',
'n1': 'b'
}
}
}
from pprint import pprint
for (k, v) in targets.items():
v = Target(v)
v.get_graph_info(group_by={})
targets[k] = v
graphs, _query = build_graphs_from_targets(targets, query)
# TODO: there should be only 1 graph, containing 3 lines, with each 2 targets per server
# i.e. something like this:
expected = {
'targets': {
{'id': ['web1.a.a', 'web1.a']},
{'id': ['web2.a.a', 'web2.a.b']},
{'id': ['web3.a.a', 'web3.b']}
}
}
print "Graphs:"
for (k, v) in graphs.items():
print "graph key"
pprint(k)
print "val:"
pprint(v)
assert expected == graphs
def test_aggregation():
preferences = DummyPrefs()
# note: uneven aggregation: we only want 1 resulting metric,
query = Query("")
query['avg_by'] = {'server': ['']}
query['sum_by'] = {'type': ['']}
targets = {
'web1.db': {
'id': 'web1.db',
'tags': {
'server': 'web1',
'type': 'db',
'n3': 'foo'
}
},
'web1.php': {
'id': 'web1.php',
'tags': {
'server': 'web1',
'type': 'php',
'n3': 'foo'
}
},
'web2.db': {
'id': 'web2.db',
'tags': {
'server': 'web2',
'type': 'db',
'n3': 'foo'
}
},
'web2.php': {
'id': 'web2.php',
'tags': {
'server': 'web2',
'type': 'php',
'n3': 'foo'
}
},
'web2.memcache': {
'id': 'web2.memcache',
'tags': {
'server': 'web2',
'type': 'memcache',
'n3': 'foo'
}
}
}
from pprint import pprint
for (k, v) in targets.items():
v = Target(v)
v.get_graph_info(group_by={})
targets[k] = v
graphs, _query = g.build_from_targets(targets, query, preferences)
# TODO: there should be only 1 graph, containing all 5 items
print "Graphs:"
for (k, v) in graphs.items():
print "graph key"
pprint(k)
print "val:"
pprint(v)
assert {} == graphs
def build_from_targets(targets, query, preferences):
graphs = {}
if not targets:
return (graphs, query)
group_by = query['group_by']
sum_by = query['sum_by']
avg_by = query['avg_by']
avg_over = query['avg_over']
# i'm gonna assume you never use second and your datapoints are stored with
# minutely resolution. later on we can use config options for this (or
# better: somehow query graphite about it)
# note, the day/week/month numbers are not technically accurate, but
# since we're doing movingAvg that's ok
averaging = {
'M': 1,
'h': 60,
'd': 60 * 24,
'w': 60 * 24 * 7,
'mo': 60 * 24 * 30
}
if avg_over is not None:
avg_over_amount = avg_over[0]
avg_over_unit = avg_over[1]
if avg_over_unit in averaging.keys():
multiplier = averaging[avg_over_unit]
query['target_modifiers'].append(
Query.graphite_function_applier('movingAverage',
avg_over_amount * multiplier))
# for each group_by bucket, make 1 graph.
# so for each graph, we have:
# the "constants": tags in the group_by
# the "variables": tags not in the group_by, which can have arbitrary
# values, or different values from a group_by tag that match the same
# bucket pattern
# go through all targets and group them into graphs:
for _target_id, target_data in sorted(targets.items()):
# FWIW. has an 'id' which timeserieswidget doesn't care about
target = Target(target_data)
target['target'] = target['id']
(graph_key, constants) = target.get_graph_info(group_by)
if graph_key not in graphs:
graph = {'from': query['from'], 'until': query['to']}
graph.update({'constants': constants, 'targets': []})
graphs[graph_key] = graph
graphs[graph_key]['targets'].append(target)
# ok so now we have a graphs dictionary with a graph for every appropriate
# combination of group_by tags, and each graph contains all targets that
# should be shown on it. but the user may have asked to aggregate certain
# targets together, by summing and/or averaging across different values of
# (a) certain tag(s). let's process the aggregations now.
if (sum_by or avg_by):
for (graph_key, graph_config) in graphs.items():
graph_config['targets_sum_candidates'] = {}
graph_config['targets_avg_candidates'] = {}
graph_config['normal_targets'] = []
# process equivalence rules, see further down.
filter_candidates = {}
for tag, buckets in sum_by.items():
# first separate the individuals from the _sum_
filter_candidates[tag] = {}
for target in graph_config['targets']:
# we can use agg_key to find out if they all have the same values
# other than this one particular key
key = target.get_agg_key({tag: buckets})
if key not in filter_candidates[tag]:
filter_candidates[tag][key] = {'individuals': []}
if target['tags'].get(tag, '') == '_sum_':
filter_candidates[tag][key]['_sum_'] = target
else:
filter_candidates[tag][key]['individuals'].append(
target)
# for all agg keys that only have the '' bucket,
# if targets are identical except that some have tag
# foo={bar,baz,0,quux, ...} and one of them has foo=_sum_ and we're
# summing by that tag, and we didn't filter on foo,
# remove all the ones except the sum one
if len(buckets) == 1 and buckets[0] == '':
if not Query.filtered_on(query, tag):
for key in filter_candidates[tag].keys():
if '_sum_' in filter_candidates[tag][key]:
for i in filter_candidates[tag][key][
'individuals']:
graph_config['targets'].remove(i)
# if we are summing, and we have a filter, and we have individual ones and a _sum_, remove the _sum_
# irrespective of buckets. note that this removes the _sum_ target without the user needing to filter it out explicitly
# this is the only place we do that, but it makes sense. we wouldn't want users to specify the _sum_ removal explicitly
# all the time, esp for multiple tag keys
if Query.filtered_on(query, tag):
for key in filter_candidates[tag].keys():
if '_sum_' in filter_candidates[tag][key]:
graph_config['targets'].remove(
filter_candidates[tag][key]['_sum_'])
for target in graph_config['targets']:
sum_id = target.get_agg_key(sum_by)
if sum_id:
if sum_id not in graph_config['targets_sum_candidates']:
graphs[graph_key]['targets_sum_candidates'][
sum_id] = []
graph_config['targets_sum_candidates'][sum_id].append(
target)
for (sum_id,
targets) in graph_config['targets_sum_candidates'].items():
if len(targets) > 1:
for candidate in targets:
graph_config['targets'].remove(candidate)
graph_config['targets'].append(
t.graphite_func_aggregate(targets, sum_by,
"sumSeries"))
for target in graph_config['targets']:
# Now that any summing is done, we look at aggregating by
# averaging because avg(foo+bar+baz) is more efficient
# than avg(foo)+avg(bar)+avg(baz)
# aggregate targets (whether those are sums or regular ones)
avg_id = target.get_agg_key(avg_by)
if avg_id:
if avg_id not in graph_config['targets_avg_candidates']:
graph_config['targets_avg_candidates'][avg_id] = []
graph_config['targets_avg_candidates'][avg_id].append(
target)
for (avg_id,
targets) in graph_config['targets_avg_candidates'].items():
if len(targets) > 1:
for candidate in targets:
graph_config['targets'].remove(candidate)
graph_config['targets'].append(
t.graphite_func_aggregate(targets, avg_by,
"averageSeries"))
# remove targets/graphs over the limit
graphs = limit_targets(graphs, query['limit_targets'])
# Apply target modifiers (like movingAverage, summarize, ...)
for (graph_key, graph_config) in graphs.items():
for target in graph_config['targets']:
for target_modifier in query['target_modifiers']:
target_modifier(target, graph_config)
# if in a graph all targets have a tag with the same value, they are
# effectively constants, so promote them. this makes the display of the
# graphs less rendundant and makes it easier to do config/preferences
# on a per-graph basis.
for (graph_key, graph_config) in graphs.items():
# get all variable tags throughout all targets in this graph
tags_seen = set()
for target in graph_config['targets']:
for tag_name in target['variables'].keys():
tags_seen.add(tag_name)
# find effective constants from those variables,
# and effective variables. (unset tag is a value too)
first_values_seen = {}
effective_variables = set() # tags for which we've seen >1 values
for target in graph_config['targets']:
for tag_name in tags_seen:
# already known that we can't promote, continue
if tag_name in effective_variables:
continue
tag_value = target['variables'].get(tag_name, None)
if tag_name not in first_values_seen:
first_values_seen[tag_name] = tag_value
elif tag_value != first_values_seen[tag_name]:
effective_variables.add(tag_name)
effective_constants = tags_seen - effective_variables
# promote the effective_constants by adjusting graph and targets:
graph_config['promoted_constants'] = {}
for tag_name in effective_constants:
graph_config['promoted_constants'][tag_name] = first_values_seen[
tag_name]
for target in graph_config['targets']:
target['variables'].pop(tag_name, None)
# now that graph config is "rich", merge in settings from preferences
constants = dict(graph_config['constants'].items() +
graph_config['promoted_constants'].items())
for graph_option in get_action_on_rules_match(
preferences.graph_options, constants):
if isinstance(graph_option, dict):
graph_config.update(graph_option)
else:
graph_config = graphs[graph_key] = graph_option(graph_config)
# but, the query may override some preferences:
override = {}
if query['statement'] == 'lines':
override['state'] = 'lines'
if query['statement'] == 'stack':
override['state'] = 'stacked'
if query['min'] is not None:
override['yaxis'] = override.get('yaxis', {})
override['yaxis'].update({'min': convert.parse_str(query['min'])})
if query['max'] is not None:
override['yaxis'] = override.get('yaxis', {})
override['yaxis'].update({'max': convert.parse_str(query['max'])})
graphs[graph_key].update(override)
# now that some constants are promoted, we can give the graph more
# unique keys based on all (original + promoted) constants. this is in
# line with the meaning of the graph ("all targets with those constant
# tags"), but more importantly: this fixes cases where some graphs
# would otherwise have the same key, even though they have a different
# set of constants, this can manifest itself on dashboard pages where
# graphs for different queries are shown.
# note that we can't just compile constants + promoted_constants,
# part of the original graph key is also set by the group by (which, by
# means of the bucket patterns doesn't always translate into constants),
# we solve this by just including the old key.
new_graphs = {}
for (graph_key, graph_config) in graphs.items():
new_key = ','.join('%s=%s' % i
for i in graph_config['promoted_constants'].items())
new_key = '%s__%s' % (graph_key, new_key)
new_graphs[new_key] = graph_config
graphs = new_graphs
return (graphs, query)
