Amazon Timestream

Amazon Timestream on fakecloud: one crate serving BOTH the Timestream Write and Timestream Query SDK clients over a shared store -- databases, tables, WriteRecords ingestion, a bounded SQL Query engine, scheduled queries, batch-load tasks, account settings, and endpoint discovery that resolves back to fakecloud -- with account-partitioned persistence.

fakecloud implements Amazon Timestream (timestream), the serverless time-series database. AWS ships it as two separate SDK clients -- aws-sdk-timestreamwrite and aws-sdk-timestreamquery -- but they belong to the same service family and, in the wire protocol, share the same awsJson1_0 target prefix Timestream_20181101.<Operation>. fakecloud serves both from a single crate over one account-partitioned store, so a database or table created through the Write client is read straight back by the Query client. All 30 distinct operations ship now (19 write + 15 query, with DescribeEndpoints, ListTagsForResource, TagResource, and UntagResource shared), backed by state that persists across restarts in persistent mode.

Every operation runs model-driven input validation first (required / length / range / enum), then real, persisted behavior -- no stubbed success responses. The declared exceptions are used throughout: ValidationException, ResourceNotFoundException, ConflictException, RejectedRecordsException, ThrottlingException, AccessDeniedException, InternalServerException, ServiceQuotaExceededException, and InvalidEndpointException.

Endpoint discovery

Timestream SDKs use endpoint discovery: the client can call DescribeEndpoints and route subsequent calls to the returned address. fakecloud's DescribeEndpoints returns an Endpoints list whose Address points back at the fakecloud host serving the request (derived from the Host header) with a large CachePeriodInMinutes (1440), so a real SDK's discovery step resolves back to fakecloud rather than to a real AWS endpoint.

The end-to-end tests drive the real aws-sdk-timestreamwrite and aws-sdk-timestreamquery clients against a live fakecloud server: create database -> create table -> WriteRecords -> Query returns the written rows. (The AWS SDK's opt-in auto-discovery mode rewrites the discovered address to https://<address>, which a plain-HTTP fakecloud server can't serve, so the tests use the default non-discovery client pinned to the fakecloud endpoint and call DescribeEndpoints explicitly to prove the host echo.)

Databases and tables

CreateDatabase mints arn:aws:timestream:<region>:<account>:database/<name> and stores the optional KmsKeyId, a TableCount, and CreationTime; DescribeDatabase / ListDatabases / UpdateDatabase / DeleteDatabase round-trip it. Deleting a database that still holds tables is a ValidationException, matching AWS.

CreateTable lives under a database (missing database -> ResourceNotFoundException), mints arn:aws:timestream:<region>:<account>:database/<db>/table/<name>, echoes back RetentionProperties (MemoryStoreRetentionPeriodInHours + MagneticStoreRetentionPeriodInDays), MagneticStoreWriteProperties, and the Schema (composite partition key), and reports status ACTIVE. ListTables filters by DatabaseName.

Ingestion

WriteRecords validates each record (dimensions, MeasureName, MeasureValue / MeasureValueType, Time / TimeUnit), merges CommonAttributes into every record, normalizes the time to nanoseconds, and stores the points in a bounded in-memory buffer per table so they become queryable. It returns RecordsIngested (Total / MemoryStore / MagneticStore). Malformed records are collected and returned as RejectedRecords inside a RejectedRecordsException (the batch is rejected atomically, as on AWS).

Query

Query runs a real, bounded SQL handler over the ingested points. PrepareQuery returns the query's ColumnInfo plus any named Parameters; CancelQuery tears down an in-flight query id. Results paginate with a round-tripping NextToken.

Honest gap -- the query SQL subset. A faithful Timestream SQL engine (the full Trino-derived time-series dialect) is out of scope. The Query handler is a genuine interpreter for the shapes SDKs, the e2e suite, and conformance exercise:

  • SELECT * FROM "database"."table" -- every stored point as a row, with AWS-shaped ColumnInfo/Datum typing: one VARCHAR column per dimension name, measure_name, a measure_value::<type> column per measure value type present, and time (TIMESTAMP).
  • SELECT COUNT(*) FROM "database"."table" -- a single BIGINT row.
  • an optional WHERE time <op> ago(<n><unit>) time filter.
  • an optional ORDER BY time [ASC|DESC].
  • an optional LIMIT <n>.

Any construct outside this subset returns a well-formed ValidationException naming the unsupported construct, rather than a wrong or silently-empty success. There is no real time-series storage engine underneath: ingested points live in a bounded buffer, not a columnar time-partitioned store.

Scheduled queries, batch load, and account settings

Scheduled queries support the full lifecycle -- CreateScheduledQuery (ARN .../scheduled-query/<name>-<id>), Describe / List / Update (state ENABLED / DISABLED) / Delete, and ExecuteScheduledQuery (records a manual run). Batch-load tasks support Create / Describe / List / Resume. DescribeAccountSettings / UpdateAccountSettings round-trip MaxQueryTCU, QueryPricingModel (BYTES_SCANNED / COMPUTE_UNITS), and the provisioned QueryCompute capacity.

Tagging

TagResource / UntagResource / ListTagsForResource are ARN-keyed and shared across databases, tables, and scheduled queries.