ms-ai-architect/skills/ms-ai-engineering/references/data-engineering/data-pipeline-orchestration.md

Data Pipeline Orchestration and Scheduling

Last updated: 2026-06-24 Status: GA Category: Data Engineering for AI Type: reference Source: https://learn.microsoft.com/fabric/data-factory/data-factory-overview

---

Innhold

• Introduksjon
• Pipeline Scheduling and Triggers
• Dependency Chains and Critical Paths
• Retry Policies and Error Handling
• Monitoring and Alerting on Pipeline Health
• SLAs and Timeliness Guarantees
• Apache Airflow i Fabric
• Referanser
• For arkitekten

Introduksjon

Datapipeline-orkestrering er ryggraden i enhver AI-plattform. Uten palitelig orkestrering kan data komme for sent, i feil rekkefølge, eller med manglende avhengigheter -- noe som forer til feil i ML-treningsjobber, utdaterte prediksjoner og upaalitelige AI-agenter. Microsoft tilbyr to hovedplattformer for orkestrering: Fabric Data Factory og Azure Data Factory, begge med pipeline-basert arbeidsflyt, triggers og overvaking.

Fabric Data Factory er den foretrukne losningen for organisasjoner som bruker Microsoft Fabric, med native integrasjon mot OneLake, Lakehouse, Warehouse og notebooks. Azure Data Factory (klassisk) gir bredere tilkoblingsmuligheter og hybrid-stotte via self-hosted integration runtime. For komplekse DAG-baserte arbeidsflyter stotter Fabric ogsa Apache Airflow-integrasjon.

For norsk offentlig sektor, der etterlevelse av SLAer og sporbarhet er kritisk, gir pipeline-orkestrering i Fabric full audit trail, automatisert feilhaandtering og mulighet for CI/CD-basert deployment av datapipelines pa tvers av miljoer.

---

Pipeline Scheduling and Triggers

Typer triggere i Fabric Data Factory

Trigger-type	Beskrivelse	Bruksomrade
Schedule	Tidsbasert med frekvens og tidsvindu	Daglige ETL-jobber, rapportoppdatering
Tumbling Window	Tidsbaserte vindu med avhengigheter	Sekvensielle batch-jobber
Event-based	Reagerer pa hendelser (ny fil, DB-endring)	Realtime-naer inntak
On-demand	Manuell kjoring	Testing, ad-hoc-jobber

Schedule Trigger-konfigurasjon

{
    "type": "ScheduleTrigger",
    "properties": {
        "description": "Daglig AI-treningsdata-oppdatering",
        "runtimeState": "Started",
        "recurrence": {
            "frequency": "Day",
            "interval": 1,
            "startTime": "2026-01-01T02:00:00Z",
            "endTime": "2027-01-01T02:00:00Z",
            "timeZone": "W. Europe Standard Time",
            "schedule": {
                "hours": [2],
                "minutes": [0]
            }
        },
        "pipelines": [
            {
                "pipelineReference": {
                    "referenceName": "IngestTrainingData",
                    "type": "PipelineReference"
                },
                "parameters": {
                    "processDate": "@trigger().scheduledTime"
                }
            }
        ]
    }
}

Event-based Trigger

{
    "type": "BlobEventsTrigger",
    "properties": {
        "description": "Trigger pa nye filer i landing zone",
        "events": ["Microsoft.Storage.BlobCreated"],
        "scope": "/subscriptions/{sub}/resourceGroups/{rg}/providers/Microsoft.Storage/storageAccounts/{sa}",
        "blobPathBeginsWith": "/landing-zone/ai-data/",
        "blobPathEndsWith": ".parquet",
        "pipelines": [
            {
                "pipelineReference": {
                    "referenceName": "ProcessNewDataFile",
                    "type": "PipelineReference"
                },
                "parameters": {
                    "fileName": "@triggerBody().fileName",
                    "folderPath": "@triggerBody().folderPath"
                }
            }
        ]
    }
}

Planlegging i Fabric UI

# Fabric pipelines kan ogsa planlegges via REST API
import requests

schedule_payload = {
    "enabled": True,
    "configuration": {
        "type": "Daily",
        "startDateTime": "2026-02-01T02:00:00.000Z",
        "endDateTime": "2026-12-31T23:59:59.000Z",
        "localTimeZoneId": "W. Europe Standard Time",
        "times": ["02:00"]
    }
}

response = requests.post(
    f"https://api.fabric.microsoft.com/v1/workspaces/{workspace_id}/items/{pipeline_id}/jobs/instances?jobType=Pipeline",
    headers=headers,
    json=schedule_payload
)

---

Dependency Chains and Critical Paths

Aktivitetsavhengigheter

Fabric Data Factory stotter fire typer avhengigheter mellom aktiviteter:

Betingelse	Beskrivelse	Bruksomrade
Succeeded	Kjor kun hvis forrige lyktes	Standard dataflyt
Failed	Kjor kun hvis forrige feilet	Feilhaandtering, alerting
Completed	Kjor uansett utfall	Opprydding, logging
Skipped	Kjor hvis forrige ble hoppet over	Betinget logikk

Kompleks avhengighetsgraf for AI-pipeline

[Ingest Raw Data]
    |-- Succeeded --> [Validate Schema]
    |                    |-- Succeeded --> [Transform Bronze->Silver]
    |                    |                    |-- Succeeded --> [Generate Features]
    |                    |                    |                    |-- Succeeded --> [Train Model]
    |                    |                    |                    |-- Failed --> [Alert: Feature Gen Failed]
    |                    |                    |-- Failed --> [Alert: Transform Failed]
    |                    |-- Failed --> [Reject and Log Invalid Data]
    |-- Failed --> [Alert: Ingestion Failed]
    |-- Completed --> [Log Pipeline Metrics]

Kritisk sti-analyse

# Beregn kritisk sti for en pipeline med flere parallelle grener
from datetime import timedelta

pipeline_activities = {
    "ingest_traffic": {"duration": timedelta(minutes=15), "depends_on": []},
    "ingest_weather": {"duration": timedelta(minutes=10), "depends_on": []},
    "ingest_road_conditions": {"duration": timedelta(minutes=12), "depends_on": []},
    "validate_traffic": {"duration": timedelta(minutes=5), "depends_on": ["ingest_traffic"]},
    "validate_weather": {"duration": timedelta(minutes=3), "depends_on": ["ingest_weather"]},
    "join_datasets": {"duration": timedelta(minutes=20), "depends_on": ["validate_traffic", "validate_weather", "ingest_road_conditions"]},
    "generate_features": {"duration": timedelta(minutes=30), "depends_on": ["join_datasets"]},
    "train_model": {"duration": timedelta(minutes=45), "depends_on": ["generate_features"]},
    "evaluate_model": {"duration": timedelta(minutes=10), "depends_on": ["train_model"]},
    "deploy_model": {"duration": timedelta(minutes=5), "depends_on": ["evaluate_model"]}
}

def find_critical_path(activities):
    """Finn den lengste stien gjennom pipeline-grafen."""
    memo = {}

    def longest_path(activity):
        if activity in memo:
            return memo[activity]

        deps = activities[activity]["depends_on"]
        if not deps:
            memo[activity] = activities[activity]["duration"]
        else:
            max_dep_time = max(longest_path(dep) for dep in deps)
            memo[activity] = max_dep_time + activities[activity]["duration"]

        return memo[activity]

    for act in activities:
        longest_path(act)

    critical = max(memo, key=memo.get)
    total_time = memo[critical]

    return total_time, memo

total, paths = find_critical_path(pipeline_activities)
print(f"Kritisk sti total tid: {total}")
# Kritisk sti: ingest_traffic -> validate_traffic -> join -> features -> train -> evaluate -> deploy
# = 15 + 5 + 20 + 30 + 45 + 10 + 5 = 130 minutter

Parallelle aktiviteter i Fabric Pipelines

{
    "name": "ParallelIngestion",
    "type": "ForEach",
    "typeProperties": {
        "isSequential": false,
        "batchCount": 5,
        "items": {
            "value": "@pipeline().parameters.dataSources",
            "type": "Expression"
        },
        "activities": [
            {
                "name": "CopyFromSource",
                "type": "Copy",
                "inputs": [{"referenceName": "@item().sourceName"}],
                "outputs": [{"referenceName": "LakehouseSink"}]
            }
        ]
    }
}

---

Retry Policies and Error Handling

Innebygde retry-policies

Parameter	Standard	Anbefalt for AI	Beskrivelse
retry	0	2-3	Antall forsok ved feil
retryIntervalInSeconds	30	60	Ventetid mellom forsok
timeout	7 dager	Varierer	Maks kjoringstid
secureInput	false	true (for tokens)	Skjul sensitive inputs

Aktivitetsniva retry

{
    "name": "FetchExternalData",
    "type": "WebActivity",
    "policy": {
        "retry": 3,
        "retryIntervalInSeconds": 60,
        "timeout": "01:00:00",
        "secureInput": false,
        "secureOutput": false
    },
    "typeProperties": {
        "url": "https://api.external-source.no/data",
        "method": "GET"
    }
}

Error Handling med kontrollflyt

{
    "activities": [
        {
            "name": "TryProcessData",
            "type": "ExecutePipeline",
            "dependsOn": [],
            "typeProperties": {
                "pipeline": {"referenceName": "ProcessDataPipeline"}
            }
        },
        {
            "name": "OnSuccess_UpdateStatus",
            "type": "SetVariable",
            "dependsOn": [
                {"activity": "TryProcessData", "dependencyConditions": ["Succeeded"]}
            ],
            "typeProperties": {
                "variableName": "pipelineStatus",
                "value": "SUCCESS"
            }
        },
        {
            "name": "OnFailure_SendAlert",
            "type": "WebActivity",
            "dependsOn": [
                {"activity": "TryProcessData", "dependencyConditions": ["Failed"]}
            ],
            "typeProperties": {
                "url": "@pipeline().parameters.alertWebhookUrl",
                "method": "POST",
                "body": {
                    "pipeline": "@pipeline().Pipeline",
                    "runId": "@pipeline().RunId",
                    "error": "@activity('TryProcessData').Error.message",
                    "timestamp": "@utcnow()"
                }
            }
        },
        {
            "name": "OnFailure_LogToTable",
            "type": "Script",
            "dependsOn": [
                {"activity": "TryProcessData", "dependencyConditions": ["Failed"]}
            ],
            "typeProperties": {
                "scriptBlockExecutionTimeout": "02:00:00",
                "scripts": [
                    {
                        "type": "NonQuery",
                        "text": "INSERT INTO dbo.pipeline_errors (pipeline_name, run_id, error_message, error_time) VALUES ('@{pipeline().Pipeline}', '@{pipeline().RunId}', '@{activity('TryProcessData').Error.message}', GETUTCDATE())"
                    }
                ]
            }
        }
    ]
}

Dead Letter Pattern for AI-data

# For feilede dataposter: flytt til dead letter-tabell i stedet for a feile hele pipeline
def process_with_dead_letter(df, transform_func, dead_letter_table):
    """
    Prosesser data med dead letter-moenster.
    Feilede rader sendes til dead letter-tabell for manuell gjennomgang.
    """
    from pyspark.sql import functions as F

    try:
        # Forsok transformasjon
        result_df = transform_func(df)
        return result_df

    except Exception as e:
        # Ved feil: forsok rad-for-rad
        success_rows = []
        error_rows = []

        for row in df.collect():
            try:
                row_df = spark.createDataFrame([row])
                transformed = transform_func(row_df)
                success_rows.append(transformed.first())
            except Exception as row_error:
                error_row = row.asDict()
                error_row["_error_message"] = str(row_error)
                error_row["_error_timestamp"] = datetime.now().isoformat()
                error_rows.append(error_row)

        # Lagre feilede rader
        if error_rows:
            error_df = spark.createDataFrame(error_rows)
            error_df.write.format("delta").mode("append") \
                .saveAsTable(dead_letter_table)

        if success_rows:
            return spark.createDataFrame(success_rows)

        return spark.createDataFrame([], df.schema)

---

Monitoring and Alerting on Pipeline Health

Fabric Monitor Hub

Fabric Monitor Hub gir enhetlig overvaking pa tvers av alle pipeline-typer:

Metrisk	Beskrivelse	Alerting-terskel
Run Status	Succeeded/Failed/In Progress	Varsle ved Failed
Duration	Kjoringstid per pipeline	Varsle ved > 2x normal
Activity Duration	Tid per aktivitet	Identifiser flaskehalser
Data Volume	Antall rader / bytes prosessert	Varsle ved 0 rader
Queue Time	Ventetid for kapasitet	Varsle ved > 5 min

REST API for pipeline-monitorering

# Hent pipeline-kjoringshistorikk
def get_pipeline_run_history(workspace_id: str, pipeline_id: str, days: int = 7):
    """Hent kjoringshistorikk for en pipeline."""
    response = requests.get(
        f"https://api.fabric.microsoft.com/v1/workspaces/{workspace_id}/items/{pipeline_id}/jobs/instances",
        headers=headers,
        params={"startDateTime": (datetime.now() - timedelta(days=days)).isoformat()}
    )

    runs = response.json()["value"]

    # Analyser
    total = len(runs)
    succeeded = sum(1 for r in runs if r["status"] == "Completed")
    failed = sum(1 for r in runs if r["status"] == "Failed")
    avg_duration = sum(r.get("durationInMs", 0) for r in runs) / max(total, 1)

    return {
        "total_runs": total,
        "success_rate": round(succeeded / max(total, 1) * 100, 1),
        "failed_count": failed,
        "avg_duration_minutes": round(avg_duration / 60000, 1)
    }

Custom Dashboard med Power BI

# Skriv pipeline-metrikker til Fabric Lakehouse for Power BI
def log_pipeline_metrics(pipeline_name: str, run_id: str, metrics: dict):
    """Logg pipeline-metrikker til overvakningstabell."""
    from pyspark.sql.types import StructType, StructField, StringType, TimestampType, LongType, DoubleType

    schema = StructType([
        StructField("pipeline_name", StringType()),
        StructField("run_id", StringType()),
        StructField("start_time", TimestampType()),
        StructField("end_time", TimestampType()),
        StructField("duration_seconds", LongType()),
        StructField("status", StringType()),
        StructField("rows_processed", LongType()),
        StructField("bytes_processed", LongType()),
        StructField("error_message", StringType()),
        StructField("sla_met", StringType())
    ])

    row = spark.createDataFrame([{
        "pipeline_name": pipeline_name,
        "run_id": run_id,
        **metrics
    }], schema)

    row.write.format("delta").mode("append") \
        .saveAsTable("lakehouse.default.pipeline_monitoring")

---

SLAs and Timeliness Guarantees

Definere pipeline-SLAer

SLA-type	Definisjon	Eksempel
Freshness SLA	Data skal vaere tilgjengelig innen X tid	"Gaarsdagens data klar for 06:00"
Completeness SLA	Alle forventede data skal vaere med	"100% av tellepunkter representert"
Quality SLA	Data skal oppfylle kvalitetskrav	"< 0.1% feilrater i features"
Availability SLA	Pipeline skal kjore X% av tiden	"99.5% tilgjengelighet"

SLA-monitorering i Fabric

# Implementer SLA-sjekk som kjorer etter pipeline
def check_pipeline_sla(pipeline_name: str, expected_completion: str, tolerance_minutes: int = 30):
    """
    Sjekk om pipeline fullforte innenfor SLA.

    Args:
        pipeline_name: Navn pa pipeline
        expected_completion: Forventet ferdigtid (HH:MM)
        tolerance_minutes: Toleranse i minutter
    """
    from datetime import datetime, time

    # Hent siste kjoring
    last_run = get_latest_pipeline_run(pipeline_name)

    if not last_run:
        return {"sla_met": False, "reason": "Ingen kjoring funnet"}

    # Parse forventet tid
    expected_time = datetime.strptime(expected_completion, "%H:%M").time()
    actual_completion = last_run["end_time"].time()

    # Beregn avvik
    expected_dt = datetime.combine(datetime.today(), expected_time)
    actual_dt = datetime.combine(datetime.today(), actual_completion)
    delay_minutes = (actual_dt - expected_dt).total_seconds() / 60

    sla_met = delay_minutes <= tolerance_minutes

    return {
        "sla_met": sla_met,
        "expected": expected_completion,
        "actual": actual_completion.strftime("%H:%M"),
        "delay_minutes": max(0, delay_minutes),
        "tolerance_minutes": tolerance_minutes,
        "status": last_run["status"]
    }

# Eksempel: Sjekk SLA for daglig AI-treningsdata
sla_result = check_pipeline_sla(
    pipeline_name="DailyAITrainingData",
    expected_completion="06:00",
    tolerance_minutes=30
)

Azure Data Factory SLA-operasjonalisering

Azure Data Factory tilbyr innebygde SLA-mekanismer for produksjonspipelines:

{
    "name": "TumblingWindowWithSLA",
    "type": "TumblingWindowTrigger",
    "properties": {
        "frequency": "Hour",
        "interval": 1,
        "startTime": "2026-01-01T00:00:00Z",
        "delay": "00:15:00",
        "maxConcurrency": 1,
        "retryPolicy": {
            "count": 3,
            "intervalInSeconds": 300
        },
        "dependsOn": [
            {
                "type": "TumblingWindowTriggerDependencyReference",
                "referenceTrigger": {
                    "referenceName": "UpstreamDataReady"
                },
                "offset": "-01:00:00",
                "size": "01:00:00"
            }
        ]
    }
}

---

Apache Airflow i Fabric

For komplekse DAG-baserte arbeidsflyter:

# Fabric stotter Apache Airflow for avansert orkestrering
# Opprett Airflow-jobb i Fabric Data Factory

from airflow import DAG
from airflow.operators.python import PythonOperator
from datetime import datetime, timedelta

default_args = {
    "owner": "ai-team",
    "depends_on_past": True,
    "email_on_failure": True,
    "email": ["ai-team@statens-ddt.no"],
    "retries": 2,
    "retry_delay": timedelta(minutes=5)
}

with DAG(
    "ai_training_pipeline",
    default_args=default_args,
    description="Daglig AI-treningspipeline",
    schedule_interval="0 2 * * *",  # Kl 02:00 daglig
    start_date=datetime(2026, 1, 1),
    catchup=False,
    tags=["ai", "training"]
) as dag:

    ingest = PythonOperator(
        task_id="ingest_raw_data",
        python_callable=ingest_from_sources
    )

    validate = PythonOperator(
        task_id="validate_data_quality",
        python_callable=run_quality_checks
    )

    transform = PythonOperator(
        task_id="transform_to_features",
        python_callable=generate_ml_features
    )

    train = PythonOperator(
        task_id="train_model",
        python_callable=train_ml_model,
        execution_timeout=timedelta(hours=2)
    )

    evaluate = PythonOperator(
        task_id="evaluate_model",
        python_callable=evaluate_model_performance
    )

    # Definer avhengigheter
    ingest >> validate >> transform >> train >> evaluate

---

Referanser

• What is Data Factory in Microsoft Fabric? -- Oversikt over Fabric Data Factory
• Pipeline overview -- Aktiviteter, scheduling og pipeline runs
• Run, schedule, or trigger a pipeline -- Trigger-typer og planlegging
• Choose a data pipeline orchestration technology -- Sammenligning av orkestreringsverktoy
• Deliver SLA for data pipelines -- SLA-operasjonalisering i ADF
• CI/CD for pipelines in Data Factory -- Deployment pipelines og Git-integrasjon
• REST API for pipelines -- Programmatisk pipeline-styring
• Create Apache Airflow jobs -- Airflow-integrasjon i Fabric

---

For arkitekten

• Bruk denne referansen naar kunder planlegger datapipeline-arkitektur for AI-arbeidsbelastninger, inkludert scheduling, avhengighetsstyring og feilhindtering.
• Fabric Data Factory er forstevalget for organisasjoner pa Fabric-plattformen. Azure Data Factory (klassisk) anbefales kun naar det trengs hybrid-stotte eller Self-Hosted IR.
• Dead letter-monsteret er kritisk for AI-pipelines: En feilende rad bor ikke stoppe hele pipeline -- send den til dead letter og fortsett. Dette sikrer at ML-modeller faar fersk data.
• SLA-monitorering bor vaere pa plass fra dag 1: Definer forventninger til ferskhet, kompletthet og kvalitet, og automatiser varsling ved brudd.
• For norsk offentlig sektor: Fremhev sporbarhet (audit trail) og CI/CD-stotte som viktige governance-funksjoner for a oppfylle krav i Forvaltningsloven og Arkivlova.