Single service comparison

Synapse exposes one workspace abstraction containing serverless SQL, Spark pools, and a dedicated SQL pool as separately managed resources. Huawei splits this across two services: DLI delivers serverless Spark, Flink, and Trino over lake data, while DWS is a separate GaussDB-backed managed warehouse. There is no single Synapse-style control plane; expect to model multiple service instances, separate IAM scopes, and independent metadata stores rather than one shared catalog and workspace.

APIs and SQL dialects diverge significantly. Synapse serverless SQL uses T-SQL over OPENJSON/external tables on ADLS; DLI uses Spark SQL/Flink SQL/Trino ANSI SQL 2003 over OBS and connected sources. DWS uses GaussDB SQL with Oracle/Teradata compatibility, not T-SQL. Synapse Pipelines correspond most directly to DataArts Studio, not to DLI or DWS, so orchestration and metadata linkage must be remapped rather than ported as-is.

Scaling and HA models differ. Synapse pools auto-scale within defined CU/DU limits and the dedicated pool is a provisioned cluster you resize. DLI uses elastic resource pools with job-level priority, and DWS scales GaussDB clusters up to thousands of nodes with multi-layer HA and read replicas. Operational responsibility shifts: DLI is serverless and patchless for engines, whereas DWS requires cluster sizing, distribution-column design, and vacuum/backup lifecycle management closer to the Synapse dedicated pool.

Start with workload classification. Serverless lake queries and Spark notebooks map to DLI; curated enterprise warehouse tables and high-concurrency BI governed by strict consistency map to DWS; pipeline orchestration and lineage map to DataArts Studio. Inventory T-SQL, external-table definitions, Spark notebooks, Power BI bindings, private endpoints, and MI dependencies, since none are directly portable and each determines the target mix.

Migrate data with CDM and DIS into OBS as the lake, then register it on DLI and load curated subsets into DWS. Recast serverless T-SQL views to DLI Spark SQL or Trino views, rework dedicated-pool T-SQL to GaussDB-compatible DWS SQL, and rewrite Spark notebook code against DLI Spark APIs. Rebuild pipelines in DataArts Studio and reattach identity via IAM plus DLI/DWS fine-grained privileges to mirror Synapse workspace roles.

Validate query results, row counts, schema constraints, and SLAs on representative workloads before cutover, since DLI is eventually-consistent lake analytics and DWS is strongly consistent. Measure DLI queue latency, DWS concurrency and skew, and re-baseline performance; do not assume Synapse throughput numbers transfer. Test failover, backup restore, and BI refresh timing.

Mind gaps and cost-model changes. There is no single T-SQL serverless engine, no native equivalent of Synapse Link for Cosmos/Dataverse, and no like-for-like Power BI Serverless SQL endpoint. Billing shifts to DLI compute-per-job, DWS cluster-spec, OBS storage, and DataArts resource billing, plus CDM/DIS transfer, so recompute TCO against peak concurrency and retention rather than per-TB scan pricing.