菜单
Apache Flink: stream processing unifié (batch = cas particulier du stream), event time, stateful computations, CEP, savepoints, exactly-once.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/JohnNuwan/EVA_CORE --skill apache-flink命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Concevoir et maintenir un watchdog auto-correcteur pour services HTTP — checks de santé, auto-restart, état persistant, rapports et pièges bash.
Serveur de messagerie sécurisé auto-hébergé (Signal-like) avec Flask + WebSocket + AES-256-GCM + pont EVA
ADAM-SENTINEL — Veilleur technologique 24h/24h. Scanne 10 domaines, cree des rapports, met a jour les skills, alerte sur les CVE et breaking changes.
基于 SOC 职业分类
正在显示 SKILL.md
| name | apache-flink |
| description | Apache Flink: stream processing unifié (batch = cas particulier du stream), event time, stateful computations, CEP, savepoints, exactly-once. |
| version | 1.0.0 |
| author | EVA & The Hive |
| license | Privée EVA St-Étienne |
| platforms | ["linux","macos","windows"] |
| metadata | {"EVA":{"tags":["flink","stream-processing","event-time","flink-sql","cep","beam","savepoint","state-backend"],"homepage":"https://flink.apache.org/","related_skills":["apache-kafka","apache-spark","data-quality-testing","data-pipelines","airflow-orchestration"]}} |
| prerequisites | {"commands":["python3","pip","java"],"pip_packages":["apache-flink","apache-flink-statefun","pyflink"]} |
Apache Flink est le moteur de traitement de flux le plus avancé, avec une philosophie radicale : le batch est un cas particulier du streaming à temps borné. Ses caractéristiques clés :
Cette compétence couvre : PyFlink (Python), Flink SQL, CEP, gestion d'état (RocksDB), sauvegarde et restauration (savepoints), déploiement sur YARN/Kubernetes, intégration Kafka, watermarking, et fenêtrage avancé.
Activez cette compétence lorsque l'utilisateur :
pip install apache-flink pyflink
Vérification :
from pyflink.datastream import StreamExecutionEnvironment
from pyflink.table import StreamTableEnvironment
env = StreamExecutionEnvironment.get_execution_environment()
env.set_parallelism(4)
print(f"Flink {env}"[:50])
from pyflink.datastream import StreamExecutionEnvironment
from pyflink.datastream.time_characteristic TimeCharacteristic
env = StreamExecutionEnvironment.get_execution_environment()
env.set_stream_time_characteristic(TimeCharacteristic.EventTime)
from pyflink.datastream.watermark import WatermarkStrategy
from pyflink.datastream.watermark import TimestampAssigner
from pyflink.common.time import Duration
class SensorTimestampAssigner(TimestampAssigner):
def extract_timestamp(self, value, record_timestamp):
import datetime
return datetime.datetime.fromisoformat(value['timestamp']).timestamp() * 1000
# Watermark : 10 secondes de tolérance au retard
watermark_strategy = WatermarkStrategy.for_bounded_out_of_orderness(
Duration.of_seconds(10)
).with_timestamp_assigner(SensorTimestampAssigner())
stream = env.from_collection([
{"machine_id": "M-001", "temperature": 87.3, "timestamp": "2026-07-22T10:00:00"},
{"machine_id": "M-001", "temperature": 88.1, "timestamp": "2026-07-22T10:00:30"},
# ... événements avec jusqu'à 10s de retard accepté
])
stream = stream.assign_timestamps_and_watermarks(watermark_strategy)
from pyflink.datastream.connectors import FlinkKafkaConsumer
from pyflink.common.serialization import SimpleStringSchema
properties = {
'bootstrap.servers': 'kafka:9092',
'group.id': 'flink-sensor-processor',
'auto.offset.reset': 'earliest',
}
kafka_source = FlinkKafkaConsumer(
topics='sensors-raw',
deserialization_schema=SimpleStringSchema(),
properties=properties
)
# Garantie exactly-once avec Kafka
kafka_source.set_start_from_earliest()
kafka_source.set_commit_offsets_on_checkpoints(True)
stream = env.add_source(kafka_source)
from pyflink.datastream import MapFunction, FlatMapFunction
from pyflink.datastream.state import ValueStateDescriptor
import json
# Map simple
class ParseSensor(MapFunction):
def map(self, value):
return json.loads(value)
parsed = stream.map(ParseSensor()).name("parse-json")
# FlatMap avec état
class TemperatureTracker(FlatMapFunction):
def __init__(self):
self.state = None
def open(self, runtime_context):
# État persistant par clé (machine_id)
desc = ValueStateDescriptor('max_temp_state', Types.DOUBLE())
self.state = runtime_context.get_state(desc)
def flat_map(self, value, collector):
max_temp_prev = self.state.value() or 0.0
if value['temperature'] > max_temp_prev + 5: # Pic de +5°C
collector.collect({
"alarme": "PIC_TEMPERATURE",
"machine_id": value['machine_id'],
"temperature": value['temperature'],
"precedent_max": max_temp_prev
})
if value['temperature'] > max_temp_prev:
self.state.update(value['temperature'])
tracked = parsed.key_by(lambda x: x['machine_id']) \
.flat_map(TemperatureTracker())
from pyflink.datastream.state import (
ValueStateDescriptor, ListStateDescriptor, MapStateDescriptor,
ReducingStateDescriptor, AggregatingStateDescriptor
)
# ListState : historique des températures
list_desc = ListStateDescriptor('temp_history', Types.DOUBLE())
temp_history = runtime_context.get_list_state(list_desc)
# MapState : configuration par capteur
map_desc = MapStateDescriptor('sensor_config', Types.STRING(), Types.DOUBLE())
sensor_config = runtime_context.get_map_state(map_desc)
# AggregatingState : moyenne glissante
# (plus efficace que de stocker toute la liste)
from pyflink.datastream.window import (
TumblingEventTimeWindows, SlidingEventTimeWindows,
EventTimeSessionWindows, GlobalWindows
)
from pyflink.common.time import Time
# Tumbling : fenêtre fixe de 5 min
tumbling_window = parsed \
.key_by(lambda x: x['machine_id']) \
.window(TumblingEventTimeWindows.of(Time.minutes(5))) \
.aggregate(TempAggregator())
# Sliding : fenêtre de 15 min, avance toutes les 5 min
sliding_window = parsed \
.key_by(lambda x: x['machine_id']) \
.window(SlidingEventTimeWindows.of(Time.minutes(15), Time.minutes(5))) \
.aggregate(TempAggregator())
# Session : fenêtre basée sur les écarts entre événements (> 2 min = nouvelle session)
session_window = parsed \
.key_by(lambda x: x['machine_id']) \
.window(EventTimeSessionWindows.with_gap(Time.minutes(2))) \
.aggregate(TempAggregator())
# Agrégateur personnalisé
class TempAggregator(AggregateFunction):
def create_accumulator(self):
return {'sum': 0.0, 'count': 0, 'min': float('inf'), 'max': -float('inf')}
def add(self, value, acc):
acc['sum'] += value['temperature']
acc['count'] += 1
acc['min'] = min(acc['min'], value['temperature'])
acc['max'] = max(acc['max'], value['temperature'])
return acc
def get_result(self, acc):
return {
'avg_temp': acc['sum'] / acc['count'],
'min_temp': acc['min'],
'max_temp': acc['max'],
'count': acc['count']
}
def merge(self, acc1, acc2):
return {
'sum': acc1['sum'] + acc2['sum'],
'count': acc1['count'] + acc2['count'],
'min': min(acc1['min'], acc2['min']),
'max': max(acc1['max'], acc2['max']),
}
from pyflink.datastream.window import TriggerResult, Trigger
from pyflink.datastream.window import TimeWindow
class EarlyFireTrigger(Trigger):
"""Déclenche un résultat après 5 événements ET au bout de 30s max"""
def on_element(self, element, timestamp, window, ctx):
count_ctx = ctx.get_partitioned_state(
ValueStateDescriptor('count', Types.INT()))
count = (count_ctx.value() or 0) + 1
count_ctx.update(count)
if count >= 5:
count_ctx.clear()
return TriggerResult.FIRE_AND_PURGE
return TriggerResult.CONTINUE
def on_processing_time(self, time, window, ctx):
return TriggerResult.FIRE
def on_event_time(self, time, window, ctx):
return TriggerResult.FIRE_AND_PURGE
def clear(self, window, ctx):
ctx.get_partitioned_state(
ValueStateDescriptor('count', Types.INT())).clear()
CREATE TABLE sensor_readings (
machine_id STRING,
temperature DOUBLE,
pression DOUBLE,
`timestamp` TIMESTAMP(3) METADATA FROM 'timestamp',
WATERMARK FOR `timestamp` AS `timestamp` - INTERVAL '10' SECOND
) WITH (
'connector' = 'kafka',
'topic' = 'sensors-raw',
'properties.bootstrap.servers' = 'kafka:9092',
'properties.group.id' = 'flink-sql-consumer',
'scan.startup.mode' = 'earliest-offset',
'format' = 'json',
'json.fail-on-missing-field' = 'true',
'json.ignore-parse-errors' = 'false'
);
-- Fenêtre tumbling de 5 minutes
SELECT
machine_id,
TUMBLE_END(`timestamp`, INTERVAL '5' MINUTE) AS fenetre_fin,
AVG(temperature) AS temp_moyenne,
MAX(temperature) AS temp_max,
COUNT(*) AS nb_mesures
FROM sensor_readings
GROUP BY
machine_id,
TUMBLE(`timestamp`, INTERVAL '5' MINUTE);
-- Fenêtre glissante (Hop)
SELECT
machine_id,
HOP_END(`timestamp`, INTERVAL '5' MINUTE, INTERVAL '15' MINUTE) AS fenetre_fin,
AVG(temperature) AS temp_moyenne
FROM sensor_readings
GROUP BY
machine_id,
HOP(`timestamp`, INTERVAL '5' MINUTE, INTERVAL '15' MINUTE);
-- Création d'une table de référence (machines)
CREATE TABLE machines (
machine_id STRING PRIMARY KEY NOT ENFORCED,
site STRING,
zone STRING
) WITH (
'connector' = 'jdbc',
'url' = 'jdbc:postgresql://postgres:5432/sensors',
'table-name' = 'machines',
'lookup.cache.max-rows' = '5000',
'lookup.cache.ttl' = '10min'
);
-- Jointure stream-table
SELECT
s.machine_id,
s.temperature,
m.site,
m.zone
FROM sensor_readings AS s
LEFT JOIN machines FOR SYSTEM_TIME AS OF s.proctime AS m
ON s.machine_id = m.machine_id;
CREATE TABLE temp_agg_delta (
machine_id STRING,
fenetre_fin TIMESTAMP(3),
temp_moyenne DOUBLE,
temp_max DOUBLE,
nb_mesures BIGINT
) WITH (
'connector' = 'delta',
'table-path' = 's3://data-lake/curated/temp_agg/',
'delta.auto-optimize.auto-compact' = 'true'
);
INSERT INTO temp_agg_delta
SELECT ... FROM sensor_readings ...;
from pyflink.datastream.cep.pattern import Pattern, AfterMatchSkipStrategy
from pyflink.datastream.cep import CEP
from pyflink.datastream.cep.functions import PatternProcessFunction
from pyflink.common.time import Time
# Pattern : augmentation rapide + pic critique dans les 2 min
pattern = Pattern.begin("mesure_stable") \
.where(lambda v: 80 <= v['temperature'] <= 90) \
.next("hausse_rapide") \
.where(lambda v: v['temperature'] > 100) \
.times_or_more(3) \
.within(Time.minutes(2))
# Sauter les patterns qui se chevauchent
skip_strategy = AfterMatchSkipStrategy.skip_past_last_event()
pattern = pattern.with_skip_strategy(skip_strategy)
# Appliquer le pattern
keyed_stream = parsed.key_by(lambda x: x['machine_id'])
pattern_stream = CEP.pattern(keyed_stream, pattern)
class AlertePatternFunction(PatternProcessFunction):
def process_match(self, match, ctx, collector):
hausse = match.get('hausse_rapide')
collector.collect({
"machine_id": hausse[-1]['machine_id'],
"type": "SURECHUFFEMENT_RAPIDE",
"derniere_temp": hausse[-1]['temperature'],
"nb_pics": len(hausse),
"timestamp": hausse[-1]['timestamp']
})
alerts = pattern_stream.process(AlertePatternFunction())
alerts.add_sink(...)
# Pattern combiné : alarme si température élevée ET pression chute
pattern = Pattern.begin("temp_elevee") \
.where(lambda v: v['temperature'] > 95) \
.followed_by("pression_basse") \
.where(lambda v: v['pression'] < 1.0) \
.within(Time.minutes(1))
from pyflink.datastream import StreamExecutionEnvironment
from pyflink.datastream.state_backend import RocksDBStateBackend
env = StreamExecutionEnvironment.get_execution_environment()
# RocksDB (recommandé pour les états > 1 Go)
rocks_backend = RocksDBStateBackend(
checkpoint_data_uri='file:///data/flink/checkpoints',
enable_incremental=True
)
env.set_state_backend(rocks_backend)
# HeapStateBackend (pour petits états, en mémoire)
# env.set_state_backend(StateBackend())
# Configuration RocksDB fine
config = rocks_backend.get_db_options()
config.set_option("rocksdb.block.cache-size", "256mb")
config.set_option("rocksdb.writebuffer.size", "64mb")
config.set_option("rocksdb.max.write.buffer.number", "4")
from pyflink.common.time import Duration
env.enable_checkpointing(60000) # Checkpoint toutes les 60s
env.get_checkpoint_config().set_checkpoint_storage_dir(
's3://flink-checkpoints/sensors-job/')
env.get_checkpoint_config().set_min_pause_between_checkpoints(30000)
env.get_checkpoint_config().set_checkpoint_timeout(Duration.of_minutes(10))
env.get_checkpoint_config().set_max_concurrent_checkpoints(1)
env.get_checkpoint_config().enable_externalized_checkpoints(
'RETAIN_ON_CANCELLATION') # Garde les checkpoints après arrêt
# Exactly-once (vs at-least-once)
env.get_checkpoint_config().set_checkpointing_mode('EXACTLY_ONCE')
# Déclencher un savepoint
bin/flink savepoint <job_id> s3://flink-savepoints/
# Redémarrer avec un nouveau JAR depuis un savepoint
bin/flink run -s s3://flink-savepoints/savepoint-xxxxx/ \
-c com.eva.SensorJob \
eva-flink-job-2.0.jar
# Arrêt avec savepoint
bin/flink stop --savepointPath s3://flink-savepoints/ <job_id>
apiVersion: flink.apache.org/v1beta1
kind: FlinkDeployment
metadata:
name: sensor-processor
spec:
image: eva/flink-sensor-job:1.0
flinkVersion: v1_20
serviceAccount: flink
jobManager:
resource:
memory: "4096m"
cpu: 2
replicas: 1
taskManager:
resource:
memory: "8192m"
cpu: 4
replicas: 4
job:
jarURI: local:///opt/flink/usrlib/eva-flink-job.jar
entryClass: com.eva.SensorJob
args: ["--kafka.bootstrap.servers", "kafka:9092"]
parallelism: 8
upgradeMode: savepoint
state: running
flinkConfiguration:
state.backend: rocksdb
state.checkpoints.dir: s3://flink-checkpoints/
state.savepoints.dir: s3://flink-savepoints/
high-availability: org.apache.flink.kubernetes.highavailability.KubernetesHaServices
high-availability.storageDir: s3://flink-ha/
# Vérifier le backpressure
# Flink Web UI → http://flink-dashboard:8081
# Jobs → Job actuel → Backpressure
# Ou via REST API
curl http://flink-jobmanager:8081/jobs/<job_id>/backpressure
from pyflink.datastream import StreamExecutionEnvironment
from pyflink.table import StreamTableEnvironment
# Batch = stream borné
env = StreamExecutionEnvironment.get_execution_environment()
t_env = StreamTableEnvironment.create(env)
# Lecture batch (fichier borné)
t_env.execute_sql("""
CREATE TABLE batch_source (
machine_id STRING,
temperature DOUBLE,
`timestamp` TIMESTAMP(3)
) WITH (
'connector' = 'filesystem',
'path' = 's3://data-lake/raw/sensors/date=2026-07-22/',
'format' = 'parquet'
)
""")
# Même SQL que le streaming
t_env.execute_sql("""
INSERT INTO temp_agg
SELECT machine_id, AVG(temperature), MAX(temperature)
FROM batch_source
GROUP BY machine_id
""").wait()
Idle sources sans watermark.
WatermarkStrategy.with_idleness(Duration.ofMinutes(1)).RocksDB trop lent.
rocksdb.block.cache-size (256-512 MB par TM). Utiliser enable_incremental=True.Checkpoints trop fréquents.
set_min_pause_between_checkpoints.Perte d'état après modification de code.
uid() sur les opérateurs stateful et setUidHash() pour garantir la compatibilité des savepoints. Nommer explicitement valueState avec des UIDs stables.Pas de gestion des late events.
sideOutputLateData() pour capturer les événements tardifs dans un flux secondaire.with_idleness() sur les sources potentiellement inactives.uid() défini sur tous les opérateurs stateful.