sdk-v3.0.0,otel-v2.0.0,testing-v1.4.0,telemetry-exporter-v0.0.2

[yaythomas]

Summary

With Durable Execution SDK 3.0 you can tag steps, callbacks and child contexts with your own subtypes, BatchResult reports every result that returned None, and plugins are safe under Lambda Managed Instances concurrency.

This coordinated major release includes SDK 3.0.0, OTel 2.0.0, Testing 1.4.0 and Telemetry Exporter 0.0.2 for Python 3.11 and later.

Upgrade Core 3 and OTel 2 together, including any Lambda layers that contain them. Core 3 rejects OTel 1.x registrations with PluginLoadError at handler initialization. Testing 1.4.0 requires Core 3. Read the 2.x to 3.0 migration guide before you upgrade.

• Core SDK and OpenTelemetry, breaking change: plugins=[...] takes factory objects with create_plugin(info), and each invocation gets a new plugin instance (894c2b1e).
• Core SDK, breaking change: Operation subtypes are plain strings, and a step, callback or child context can set a custom sub_type (0783da2a).
• Core SDK, breaking change: BatchResult.get_results() keeps a branch that returned None, and get_errors() keeps every failed item (e567ae4f).
• Core SDK, new feature: Plugin hooks no longer run on the checkpoint thread, so a slow hook no longer delays checkpoints (363db885).
• Testing, compatibility: Testing 1.4.0 requires Core 3.0.0 and reads subtypes as strings (0783da2a).

---

aws-durable-execution-sdk-python v3.0.0

Breaking Changes

The migration guide shows the code change for each item.

• plugins takes factory objects. A factory has a synchronous create_plugin(info: InvocationStartInfo) method that returns a new DurableInstrumentationPlugin. The SDK calls it once per invocation (894c2b1e).
• DurableInstrumentationPluginProvider is removed. An entry point in the aws_durable_execution.plugins group must resolve to a factory object. A plugin instance, a bare callable or an uninstantiated factory class in plugins raises PluginLoadError at handler initialization (894c2b1e).
• Context variables that a failed on_invocation_start set stay in place. Later hooks and handler code see them (894c2b1e).
• Every subtype field is str | None: Operation.sub_type, OperationUpdate.sub_type, ChildConfig.sub_type and OperationInfo.sub_type. Replace .value reads with the value, and is comparisons with == (0783da2a).
• OperationUpdate.create_context_start, create_context_succeed and create_context_fail lose their sub_type parameter. The operation identifier carries it. aws_durable_execution_sdk_python.plugin no longer imports OperationSubType. Import it from aws_durable_execution_sdk_python.lambda_service (0783da2a).
• BatchResult.succeeded() and get_results() include a branch that returned None. So sum(batch.get_results()) can raise TypeError where it used to work (e567ae4f).
• BatchResult.failed() and get_errors() return every failed item, one entry each. A failed child context whose checkpoint carries no error is recorded with the ChildContextError that replaying it raises, so throw_if_error() raises for it. In 2.x the item was dropped and throw_if_error() returned without raising (e567ae4f).

Features

• StepConfig and CallbackConfig take a keyword-only sub_type. ChildConfig.sub_type accepts any string. None and "" select the default (0783da2a).
• The state loader keeps any subtype string the service stored. In 2.x Operation.from_dict raised ValueError on a value outside OperationSubType (0783da2a).
• A subtype outside 1 to 32 characters from A-Z, a-z, 0-9, - and _ raises ValidationError before any checkpoint is sent. Replay compares the subtype with the checkpoint. So changing a subtype breaks executions already in flight (0783da2a).
• An operation end hook runs on the thread that observes the result, before the operation returns. on_operation_change runs on its own thread and merges notifications that arrive while the plugin is still in the hook (363db885).

Chores and maintenance

• The user agent detects a runtime-bundled SDK from a +bundled version label instead of the install path. The reported value keeps the form <version>-bundled (7aec4442).
• Seven plugin payload fields are no longer marked experimental: OperationInfo.result and error, InvocationInfo.execution_input and operations, InvocationStartInfo.updated_operations, and InvocationEndInfo.error and execution_result (8bb88f56).

---

aws-durable-execution-sdk-python-otel v2.0.0

Breaking Changes

The migration guide shows the new registrations.

• Register InvocationOtelPluginFactory or ExecutionOtelPluginFactory. InvocationOtelPlugin and ExecutionOtelPlugin are no longer valid entries in plugins (894c2b1e).
• Requires Core >=3.0.0,<4. OTel 2 cannot use the Core 2 plugin contract. DURABLE_EXECUTION_PLUGINS=otel-invocation and otel-execution keep their names but must resolve to OTel 2 (894c2b1e).

Features

• The durable.operation.subtype span attribute carries a custom subtype set on a step, callback or child context (0783da2a).

Bug Fixes

• Each invocation gets its own plugin instance. So overlapping invocations on Lambda Managed Instances no longer overwrite each other's span state (894c2b1e).

---

aws-durable-execution-sdk-python-testing v1.4.0

Breaking Changes

• Requires Core >=3.0.0. A project pinned to Core 2 cannot install it (0783da2a).
• Operation.sub_type is str | None. Code that reads .value or compares with is must compare with == (0783da2a).

Features

• The local runner accepts and records a custom subtype. 1.3.0 raised InvalidParameterValueException for it. The runner does not yet reject an invalid subtype, which the service rejects. So a subtype that passes locally can still fail in the cloud. Tracked in issue [Testing]: Local checkpoint server accepts a SubType the service rejects #836 (0783da2a).

---

aws-durable-execution-sdk-python-telemetry-exporter v0.0.2

Breaking Changes

• durable_execution_telemetry(config) returns a plugin factory. Pass its result directly to plugins. Do not wrap or cache a plugin it creates (894c2b1e).
• Requires Core >=3.0.0,<4. Version 0.0.1 uses the Core 2 plugin contract and cannot run on Core 3 (894c2b1e).

Bug Fixes

• Record timestamps use YYYY-MM-DDTHH:MM:SS.sssZ with exactly 3 fraction digits. Before, Python wrote 6 digits, or none when microseconds were 0, so text order did not match time order. emittedAt strictly increases within one invocation. So two records built in the same millisecond no longer share an SQS FIFO deduplication ID or a default DynamoDB sort key (102dd62e).
• LambdaLogExporter prints each record inside the Lambda JSON log envelope {"timestamp","level","message"}. max_record_size_bytes caps the printed line (f870be53).
• An opted-in operation result written by the default serializer is decoded to plain JSON. Integers beyond 2**53 - 1 are written as strings. DynamoDBExporter writes a number it cannot store as a string instead of failing the item (f870be53).
• Every error is written with the keys name and message, in that order. A name or message longer than 8,192 code points is cut and ends with … (f870be53).
• Every zero Decimal is written as 0. Before, Decimal("-0E+20") was written as -0.0 (b4a1d711).
• Overlapping invocations no longer lose records through shared plugin state (894c2b1e).

Full Changelog: sdk-v2.1.0,otel-v1.1.0,testing-v1.3.0...sdk-v3.0.0,otel-v2.0.0,testing-v1.4.0,telemetry-exporter-v0.0.2

---

This discussion was created from the release sdk-v3.0.0,otel-v2.0.0,testing-v1.4.0,telemetry-exporter-v0.0.2.