You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
[Worker] Start status is sent before launching the JobRunner process.
[Core] Dependency debug logs distinguish function modules from input-data modules.
[Azure Functions] Default runtime memory is now 512 MB.
[Core] A storage or monitoring service at a loopback address is refused for serverless workers, which cannot reach it; standalone and Singularity workers get a warning.
[Core] wait() timeouts no longer arm a SIGALRM: they are checked between polls, work from any thread and leave the caller's signal handler and alarm alone.
[Core] An obj_chunk_size no longer than obj_newline is raised to one byte more.
[Storage] Glob patterns in the input objects work on every storage backend: a key with * is a pattern, as it already was on AWS S3 and IBM COS, also from a Windows client.
[CLI] Times are shown in the local time zone without pytz and tzlocal, which are no longer used.
[Standalone] lithops worker list shows the last error of each worker and how many of its processes are busy.
[Knative] Updated to current Knative Serving (tested with 1.23): one container per call in flight with the current autoscaling annotations, kubecfg_context, namespace and runtime_arch keys, the Contour networking layer, cluster-local calls from inside the cluster, a deploy that updates an existing service and reports why its revision fails, and a default runtime that runs as a non-root user.
Fixed
[Code Engine] clean --all permanently deletes the project so its name can be reused.
[GCP Compute Engine] Cleanup tolerates VMs already deleted by another process.
[GCP] Compute Engine infers the project from credentials when omitted; Functions and Cloud Run honour explicit project IDs.
[GCP Functions] Cleaning runtimes works when service configuration is missing.
[Azure VMs] Fixed network request serialization with newer SDKs and enabled internet access for workers on newly created subnets.
[Runtime] Podman logins include the Docker Hub registry explicitly, so runtime builds can authenticate without a default registry in registries.conf.
[Core] get_result() with no futures returns []. call_async() chains a list or tuple of futures as one call. RetryingFunctionExecutor.wait() shares one timeout across retry rounds and, when it expires, leaves the calls running instead of cancelling them.
[Core] Chaining, cleanup, invocation errors, generator inputs, empty fs, wait() with ANY_COMPLETED, argument binding, and RetryingFunctionExecutor retries and percentage waits; a failed object, or chunk of one, is retried on its own.
[Worker] Partitioned object streams (lost, duplicated or cut rows, rows longer than 128 KB, multi-byte obj_newline, readline(), iteration), module discovery of nested functions, a parameter named future, and extra_env restoring the previous environment.
[Localhost] An absolute interpreter path with any character, such as @ or a space, is accepted as runtime; storage listing is a literal prefix match including dotfiles, and open-ended byte ranges are supported.
[Storage] CloudFileProxy path handling (isfile, listdir, exists, leading /), double close(), and r+ mode writing back; Swift generate_bucket_name, Aliyun and Oracle listings past 1,000 objects, and open-ended byte ranges on GCS, Azure and Aliyun without modifying the caller's extra_get_args; glob patterns outside the bucket root on AWS S3 and IBM COS; Swift listings return Key and Size like the other backends and go past the first page. A missing bucket from an S3-compatible listing is StorageNoSuchKeyError. Swift uploads and non-404 reads raise, Swift has head_bucket, and an empty error page is not the end of a listing. Oracle, Azure and GCP no longer report an arbitrary failure as a missing key or bucket. CloudFileProxy.exists() ignores Lithops-internal keys. Swift logs in to current Keystone: a project is scoped by id or by name and domain, the domain defaults to Default, and the catalog URL is used as published, including https. A container that does not exist is created.
[Serverless] Oracle Functions reports a missing oracle_f section, an Aliyun zip runtime builds without Docker, and Kubernetes no longer waits for ever on a runtime image it cannot pull or a master pod that does not start. The Kubernetes registry secret uses the Docker config keys. A Knative, GCP Functions or Cloud Run deploy that cannot read runtime metadata deletes the service or function it created, and GCP Functions reads that metadata once. An unknown Knative kubecfg_context names the contexts that exist. Waiting for a GCP function or Cloud Run service to be deleted treats only HTTP 404 as gone, and a failed Cloud Run delete is reported.
[Standalone] A failed init() on AWS EC2, GCP Compute Engine and IBM VPC now deletes every resource it created, including those whose tagging, attaching or rules failed, and IBM VPC floating IPs allocated by that init(); IBM VPC no longer uses the deprecated code of SDK errors; the vm backend loads its documented configuration, and Azure request_spot_instances creates Spot worker VMs (now False by default: B-series sizes, like the default worker, do not support Spot). AWS EC2 raises a clear error when spot price history is empty.
[Monitoring] Worker tokens of jobs dropped by wait(), replaced monitors and restarted invokers are handed back once, a held __end__ is no longer replaced by an __init__, and the workers of a job that owes no token are forgotten instead of kept for as long as the job monitor runs; the shared RabbitMQ connection of a worker is closed at exit.
[Concurrent] wait() and as_completed() accept generators.
[Core] An executor that fails to be created no longer logs a misleading job_monitor error at exit.
[Core] The FaaS invoker no longer runs up to two workers beyond max_workers. A wait() that times out drops the invocations still queued, the same as any other failed wait. obj_chunk_number splits an object into exactly that many chunks, or fewer non-empty ones when it is smaller. URL inputs are sized following redirects, an HTTP error raises, and only Accept-Ranges: bytes enables partitioning. A local input path that does not exist raises FileNotFoundError, and local inputs with no files raise like an empty bucket. A mode section with no body in the config file no longer crashes config loading.
[CLI] lithops runtime update stores runtime_timeout in the runtime metadata, which every later job needed.
[Standalone] Cleaning AWS EC2 workers reads every page of instances, so workers past the first 1,000 are stopped too. Cleaning IBM VPC workers, choosing its image and reusing a free floating IP read every page, which stops at 50, as do the VPC, subnet, gateway and SSH key lookups. Cleaning GCP Compute Engine workers reads every page of instances, which stops at 500. get_result() on one job no longer cancels the other jobs of the executor, and create mode stops its workers only once none of its jobs is left. Worker VMs whose setup fails, and the workers of a job whose master cannot be set up, are stopped instead of left running. Requests to the master from inside a worker send their body, and an error from the master raises. They wait up to 60 seconds for an answer instead of 1, and as long as the master needs to pull the runtime for its metadata; a POST no longer waits for ever. The vm backend runs jobs: it reads ssh_username and provides what the master and its budget keeper need. Azure VMs are deallocated when stopped, so they are no longer billed, the home directory follows ssh_username, and delete_on_dismantle defaults to False as documented.
[AWS Lambda] Invocations retry connection errors a few times, re-signing each attempt, instead of retrying any error for ever. lithops runtime update updates the code, configuration and tags of a deployed function, and the default runtime gets the user_tags too. env_vars and user_tags accept a {NAME: VALUE} mapping or a list of {name, value}. Layers with more than 50 versions are deleted. Runtimes and layers past the first page of the listing are found by lithops runtime list and clean, and deleting a runtime whose function is already gone works. A container runtime is looked up by its image tag, so it is found in a repository with more than 100 images.
[AWS Batch] Maps over more than 20 calls no longer fail on a fractional max_workers. Job definitions use the requested runtime_memory, checked against the Fargate combinations, and their names carry the requested version, so lithops runtime update and delete find them, also past the first page of job definitions. Deleting a runtime whose job definition is already gone works. Job queues and compute environments past the first page are found, and lithops runtime update registers a new job definition revision instead of keeping the CPU, roles and image from the first deploy.
[OpenWhisk] An invoke rejected with 401 or 404 raises instead of being reported as an activation id, which made wait() hang. A failure to read runtime metadata raises that error instead of a TypeError. Python 3.13 and 3.14 clients get a default runtime, built from the runtime Dockerfile with --build-arg PYTHON_VERSION. The client speaks to current OpenWhisk (tested with openwhisk/standalone:nightly): an http:// endpoint is used as configured, actions and packages past the first page are listed, a package that already exists is reused, deleting a missing action works, and waiting for runtime metadata follows the activation instead of invoking again for ever. A non-blocking invocation opens its own connection and closes it, so the invoker threads do not queue on the requests connection pool. lithops runtime build builds for the machine it runs on; --platform linux/amd64 selects amd64.
[Code Engine] Projects, job definitions and config maps past the first page are listed. clear() deletes every job run and config map of the executor, a leftover config map is replaced instead of failing every deploy, and waiting for a job run or a job definition deletion times out.
[Singularity] Runtime metadata that never arrives raises the intended timeout error.
[Kubernetes] Worker memory is compared in megabytes, so a Mi request is the same size as the memory it is checked against.
[Azure Functions] Fixed app lookup and CLI error reporting, invocation and metadata timeouts, metadata POST requests, per-memory deployments, and runtime lookup by name.
[Azure Container Apps] Added deployment progress logs and cleanup on interrupted creation. Fixed authentication error handling, runtime lookup by image, and numeric memory values.
[Aliyun FC] The synchronous metadata call waits for runtime_timeout. An asynchronous invoke only waits to be accepted. Retries while reading runtime metadata resend the payload instead of an empty body.
[Oracle Functions] Calls run on the function of their own runtime and memory, not on the function of the job started last. Functions and applications past the first page are found by lithops runtime list, delete and clean.
[GCP Functions] [GCP Cloud Run] An HTTPS invocation waits for runtime_timeout instead of giving up after 120 seconds, which is while the call is still running. Functions and services past the first page are found by lithops runtime list and clean.
[Knative] Calls longer than 5 minutes are no longer cut off by Knative, a load balancer ingress is reached at its own port and may be a hostname, https service URLs work, a failed call is reported instead of being invoked again for ever, and a service that never gets ready raises an error.