feat(destination): add Hotdata managed-database destination

[eddietejeda]

Adds Hotdata as a first-party dlt destination:

dlt.destinations.hotdata

Hotdata is a managed-database service that accepts Parquet uploads via HTTP API.

This destination implements JobClientBase and WithStateSync, the same interface as other non-SQL destinations, so it works with existing pipelines without code changes.

Feature Support

Feature	Status	Notes
replace	✅	Full table re-upload
append	✅	Permissive concat; schema drift handled
merge / upsert	✅	Client-side upsert by primary key
insert-only	✅	Insert when not matched; existing rows untouched
truncate-and-insert	✅	Declared replace strategy
Table nesting	✅	Unlimited (1000), configurable per destination
dlt metadata columns	✅	_dlt_id, _dlt_load_id, _dlt_parent_id, etc. preserved
Pipeline state sync	✅	WithStateSync; survives across runs
Schema versioning	✅	Stored in _dlt_version managed table
Load tracking	✅	Stored in _dlt_loads managed table
Auto-create database	✅	Configurable via create_database_if_missing
Schema evolution	✅	Delete and recreate DB with union of old and new tables
Retry / backoff	✅	Configurable retries, exponential backoff capped at 30 seconds
Error classification	✅	Transient 408, 409, 425, 429, 5xx vs terminal
Parallelism strategy	✅	Table-sequential default, configurable
Max table nesting	✅	Configurable per destination instance
Identifier normalisation	✅	snake_case convention: [a-z0-9_]; nested tables as parent__child
Dataset read API	❌	Requires SqlJobClientBase
SCD2	❌	Requires server-side SQL
Staging area	❌	No Hotdata staging concept
Type mapper	❌	Not needed; Parquet carries its own types
Clone table	❌	No API endpoint

New Files

dlt/destinations/impl/hotdata/
├── __init__.py
├── _api_client.py       # Hotdata SDK wrapper with retry logic
├── configuration.py     # HotdataCredentials + HotdataClientConfiguration
├── contracts.py         # Identifier normalisation, TableContract
├── errors.py            # Error classification: transient vs terminal
├── factory.py           # hotdata(Destination[...]) + capabilities
├── hotdata.py           # HotdataClient + HotdataLoadJob
├── merge.py             # combine_tables for all write dispositions
└── parquet.py           # Arrow → Parquet writer

tests/load/hotdata/
└── test_hotdata_client.py   # 44 unit tests

Related Issues

N/A — new destination.

Additional Context

• Merge is executed client-side: fetch existing data, combine in Arrow, then re-upload. Server-side merge is planned for the next Hotdata API release and will eliminate the need for the fetch step.
• loader_parallelism_strategy defaults to table-sequential to prevent concurrent read-modify-write races on the same table.
• The Hotdata API requires tables to be declared at database creation time. When a new table appears mid-pipeline, the destination deletes and recreates the database with the union of existing and new declared tables, preserving all data.