Huawei DWS, MRS and DLI migration architecture

Four target designs

Huawei Cloud services described from their documentation pages, STACKIT options from the Dremio documentation, the Terraform provider documentation and the run reasons, all read on 4 November 2026. No benchmark or price is implied.

A warehouse, a Hadoop cluster and a queue, assessed

A small Huawei Cloud data estate with eight resources, written for this page and assessed by StackToIT on 4 November 2026: DWS cluster with three nodes, MRS cluster with Hadoop, Spark and Hive, a DLI queue of 16 CU, a DLI database, an OBS bucket and the network. Left: three blocks of the input. Right: the result the run recorded for them, an excerpt of assessment.json with the values as the run wrote them (the target field is left out).

resource "huaweicloud_dws_cluster" "warehouse" {
  name              = "lake-dws"
  version           = "8.2.1"
  node_type         = "dws.m3.xlarge"
  number_of_node    = 3
  availability_zone = "eu-west-101a"
  user_name         = "dbadmin"
  user_pwd          = var.dws_password
  vpc_id            = huaweicloud_vpc.main.id
  network_id        = huaweicloud_vpc_subnet.app.id
  security_group_id = huaweicloud_networking_secgroup.app.id

  volume {
    type     = "SSD"
    capacity = 300
  }
}

resource "huaweicloud_mrs_cluster" "batch" {
  cluster_name         = "lake-mrs"
  cluster_version      = "MRS 2.1.0"
  cluster_type         = 0
  available_zone_id    = "ae04cf9d61544df3806a3feeb401b204"
  billing_type         = 12
  master_node_num      = 2
  core_node_num        = 3
  master_node_size     = "c3.4xlarge.2.linux.bigdata"
  core_node_size       = "c3.xlarge.4.linux.bigdata"
  volume_type          = "SAS"
  volume_size          = 100
  safe_mode            = 0
  cluster_admin_secret = var.mrs_admin_secret
  vpc_id               = huaweicloud_vpc.main.id
  subnet_id            = huaweicloud_vpc_subnet.app.id

  component_list {
    component_name = "Hadoop"
  }
  component_list {
    component_name = "Spark"
  }
  component_list {
    component_name = "Hive"
  }
}

resource "huaweicloud_dli_queue" "etl" {
  name     = "lake-etl"
  cu_count = 16
}

[
  {
    "id": "huaweicloud_dws_cluster.warehouse",
    "result": {
      "class": "REDESIGN",
      "confidence": "LOW",
      "reason": "Data Warehouse Service has no managed counterpart. The data and queries move to a redesigned lakehouse or warehouse."
    }
  },
  {
    "id": "huaweicloud_mrs_cluster.batch",
    "result": {
      "class": "REDESIGN",
      "confidence": "LOW",
      "reason": "MapReduce Service and Data Lake Insight have no managed counterpart. Jobs move to self-managed engines on SKE, or to a redesigned lakehouse."
    }
  },
  {
    "id": "huaweicloud_dli_queue.etl",
    "result": {
      "class": "REDESIGN",
      "confidence": "MEDIUM",
      "reason": "MapReduce Service and Data Lake Insight have no managed counterpart. Jobs move to self-managed engines on SKE, or to a redesigned lakehouse."
    }
  }
]

• REDESIGN, with the reason

  Data Warehouse Service has no managed counterpart; MapReduce Service and Data Lake Insight have none either. Each reason names the redesign rather than a copy.

• LOW confidence twice

  The DWS and MRS units are LOW, the DLI queue MEDIUM. Confidence says how sure the rule is about this unit’s configuration, not how hard the migration is.

• Critical settings left to you

  Twelve for DWS, sixteen for MRS and two for the DLI queue are listed in the architect report for your team, among them node type and node count for DWS and the administrator secret for MRS.

How the eight units come out

Classes, confidence and waves as the run decided them. The DWS and MRS units wait for the network in waves 1 and 2, but their data and jobs are a separate plan.

What the data file lists, and what it misses

Left: the stateful table of migration/data-movement.md for this run. It lists only the OBS bucket: the DWS tables, the HDFS data of the MRS cluster and the DLI database are not there, so add them to your own data plan. Right: a second, small run on the two newer resource types that the provider recommends or offers for MapReduce and for DLI compute pools. That probe input, written for this page, sets the attributes and node blocks the 1.100.0 provider schema marks as required; the result depends only on the resource type.

## Stateful source units

| Unit | Source service | Target service | Application | Mechanism | State | Data class | RPO | Downtime tolerance |
|---|---|---|---|---|---|---|---|---|
| huaweicloud_obs_bucket.raw | huaweicloud_obs_bucket | STACKIT Object Storage |  | UNKNOWN | DECISION_REQUIRED | NOT_PROVIDED | NOT_PROVIDED | NOT_PROVIDED |

[
  {
    "id": "huaweicloud_mapreduce_cluster.batch",
    "result": {
      "class": "MANUAL_REVIEW",
      "confidence": "LOW",
      "reason": "huaweicloud_mapreduce_cluster is not in the knowledge base catalogue; an architect has to decide the target. Nothing of this resource is assessed or drafted (unknown source type)."
    }
  },
  {
    "id": "huaweicloud_dli_elastic_resource_pool.pool",
    "result": {
      "class": "MANUAL_REVIEW",
      "confidence": "LOW",
      "reason": "huaweicloud_dli_elastic_resource_pool is not in the knowledge base catalogue; an architect has to decide the target. Nothing of this resource is assessed or drafted (unknown source type)."
    }
  }
]

• Only the bucket is listed

  The data file lists the OBS bucket as the stateful unit. The tables inside DWS, the HDFS content of MRS and the DLI database are data too, and the run does not list them. Inventory them by hand.

• Resource names matter

  The provider documents huaweicloud_mrs_cluster as deprecated and recommends huaweicloud_mapreduce_cluster. In this build the older name is REDESIGN, while mapreduce_cluster and dli_elastic_resource_pool are unknown types and come out MANUAL_REVIEW with LOW confidence. Read the unknown ones as the same redesign.

How to decide, workload by workload

1. 01

   1. Move the storage first

   Data in OBS becomes data in STACKIT Object Storage, a RECONFIGURE unit with an S3 API. Copy it, for example with rclone. Any lakehouse design needs the files there anyway.

2. 02

   2. Inventory what Terraform cannot show

   List schemas, tables, row counts, query logs, scheduled jobs, HDFS paths and DLI databases. The three services hold different things: MRS is a Hadoop platform with Spark, Hive, Flink and Kafka components, DLI is a serverless Spark and Flink SQL service, DWS is an analytical SQL database whose storage-compute decoupled architecture keeps user data on OBS.

3. 03

   3. Sort workloads by shape

   Interactive SQL, scheduled batch, streaming and ad hoc serverless queries do not need the same target. Streaming and Spark batch point to design B, warehouse SQL to A or C.

4. 04

   4. Prototype the hardest query

   Take the slowest or most dialect-specific query of each workload and run it on the candidate engine with a real extract. Dialect and file-format surprises show up here, not in a diagram.

5. 05

   5. Plan copy and cutover per workload

   Choose the data mechanism, the window and the order in which reporting tools repoint. The run sets no mechanism and gives no duration.

What the run does and does not decide

• It decides the class

  REDESIGN for DWS, MRS and the DLI queue, with the reason in assessment.json.

• It names candidates

  A lakehouse on Object Storage with STACKIT Dremio, or a self-managed warehouse. The Dremio resource is an opt-in provider feature that may still change, so the run records that no generally available target exists yet, as a blocker, and leaves class and confidence unchanged.

• It does not size the target

  Node counts, volumes and queue capacity are inputs for your design. The architect report lists twelve critical settings for the DWS cluster, sixteen for the MRS cluster and two for the DLI queue, and maps none of them.

• It does not see the data

  Sizes, change rates and downtime tolerance are not in the Terraform. They show as not known in the data file.

• It does not pick the region

  The provider documents Dremio for eu01 only, and the run records region availability of these targets as not established. Settle the region before the design.

• It does not move anything

  StackToIT never moves data or runs jobs. Every step is a manual step with an owner.

Questions about analytics workloads

Sources

• Huawei Cloud MapReduce Service: overview
• Huawei Cloud Data Lake Insight: introduction
• What Is DWS? (Huawei Cloud Data Warehouse Service documentation)
• STACKIT Dremio documentation
• stackit_dremio_instance (STACKIT Terraform provider 0.121.0)
• STACKIT Cloud Framework: Migration Framework