Alibaba Cloud Container Registry is now supported for the docker runtime attribute, and the previous dockerTag
runtime attribute continues to be available for Alibaba Cloud OSS Registry.
Cromwell now supports Call caching when using the BCS backend.
Globs can be used to define outputs for BCS backend.
Alibaba Cloud NAS is now supported for the mounts runtime attribute.
In some cases PAPI v2 will report the preemption of a VM in a way that differs from PAPI v1. This novel means of reporting preemption was not recognized by Cromwell's PAPI v2 backend and would result in preemptions being miscategorized as call failures. Cromwell's PAPI v2 backend will now handle this type of preemption.
Cromwell now allows PAPIV2 jobs to run on a specific subnetwork inside a private network by adding the subnetwork key
subnetwork-label-key inside virtual-private-cloud in backend configuration. More info here.
Cromwell's CALL_CACHING_HASH_ENTRY primary key has been refactored to use a BIGINT datatype in place of the previous
INT datatype. Cromwell will not be usable during the time the Liquibase migration for this refactor is running.
In the Google Cloud SQL with SSD environment this migration runs at a rate of approximately 100,000 CALL_CACHING_HASH_ENTRY
rows per second. In deployments with millions or billions of CALL_CACHING_HASH_ENTRY rows the migration may require
a significant amount of downtime so please plan accordingly. The following SQL could be used to estimate the number of
rows in this table:
select max(CALL_CACHING_HASH_ENTRY_ID) from CALL_CACHING_HASH_ENTRY
Cromwell now supports sending metrics to Google's Stackdriver API. Learn more on how to configure here.
Cromwell now allows a user to specify BigQuery jobs when using the PAPIv2 backend
There is a small change in StatsD's configuration path. Originally, the path to the config was services.Instrumentation.config.statsd
which now has been updated to services.Instrumentation.config. More info on its configuration can be found
here.
A new experimental feature, the cached-copy localization strategy is available for the shared filesystem.
More information can be found in the documentation on localization.
Yaml parsing now checks for cycles, and limits the maximum number of parsed nodes to a configurable value. It also limits the nesting depth of sequences and mappings. See the documentation on configuring YAML for more information.
- It is now possible to use
includeKeyandexcludeKeyat the same time. If so, the metadata key must match theincludeKeyand not match theexcludeKeyto be included. - It is now possible to use "
calls" as one of yourexcludeKeys, to request that only workflow metadata gets returned.
Cromwell now supports PostgreSQL (version 9.6 or higher, with the Large Object extension installed) as a database backend. See here for instructions for configuring the database connection.
The /describe endpoint now differentiates between an invalid workflow and a valid workflow with invalid inputs.
Specifically, the new validWorkflow key indicates whether the workflow file is valid by itself. If inputs are provided, they are not considered when calculating this field; if inputs are not provided, the value is identical to valid.
- Virtual private networks can now be configured. See the section below for details.
The timeout on Cromwell's requests to PAPIv2 can now be configured. See the sample PAPIv2.conf for more documentation:
backend {
providers {
PAPIv2 {
config {
batch-requests {
timeouts {
read = 10 seconds
connect = 10 seconds
}
}
}
}
}
}
Cromwell now allows PAPIV2 jobs to run on a private network by adding the network name inside virtual-private-cloud in backend configuration.
More info here.
Now includes background job status polling to hopefully reduce the incidence of 'HTTP 429' errors for large workflows.
- It is now possible to supply custom
google-labelsin workflow options.
It is now possible to use WDL disk attributes with the following formats on AWS.
disks: "local-disk 20 SSD"
disks: "/some/mnt 20 SSD"
Because Cromwell's AWS backend auto-sizes disks, the size specification is simply discarded.
In previous versions of Cromwell, times were converted to strings using the default Java formatter which generates a variety of ISO-8601 formats. String conversions also retained whatever server time zone generated that specific time instance.
Going forward, times stored in Cromwell metadata, and later returned via the HTTP endpoint, are now converted to UTC then formatted with exactly three digits of milliseconds.
For example:
2017-01-19T12:34:56-04:00will now be formatted as2017-01-19T16:34:56.000Z
This change only affects newly formatted dates. Older dates already formatted and stored by previous versions of
Cromwell will not be updated however they will still return a
valid ISO-8601 format. The older format may be in various non-UTC time zones,
and may or may not include microseconds or even nanoseconds, for example 2017-01-19T12:34:56.123456789-04:00.
When a Cromwell instance is unable to write heartbeats for some period of time it will automatically shut down. For more information see the docs on configuring Workflow Hearbeats.
NOTE: In the remote chance that the system.workflow-heartbeats.ttl has been configured to be less than 5 minutes
then the new configuration value system.workflow-heartbeats.write-failure-shutdown-duration must also be explicitly
set less than the ttl.
The runtime attribute nvidia-driver-version was previously allowed only as a default runtime attribute in configuration.
Because WDL does not allow attribute names to contain - characters, this has been changed to nvidiaDriverVersion.
This field is now accepted within WDL files as well as within the configuration file.
All backends can now emit slow job warnings after a configurable time running. NB This example shows how to configure this setting for the PAPIv2 backend:
# Emit a warning if jobs last longer than this amount of time. This might indicate that something got stuck.
backend {
providers {
PAPIv2 {
config {
slow-job-warning-time: 24 hours
}
}
}
}
- The
gpuTypeattribute is no longer validated against a whitelist at workflow submission time. Instead, validation now happens at runtime. This allows any valid accelerator to be used. - The
nvidiaDriverVersionattribute is now available in WDLruntimesections. The default continues to be390.46which applies if and only if GPUs are being used. - A default
gpuType("nvidia-tesla-k80") will now be applied ifgpuCountis specified butgpuTypeis not. - Similarly, a default
gpuCount(1) will be applied ifgpuTypeis specified butcpuCountis not.
An error will be thrown on startup when the system.workflow-heartbeats.heartbeat-interval is not less than the
system.workflow-heartbeats.ttl.
When set, the configuration value of system.cromwell_id will be prepended to StatsD metrics. More info here.
The HealthMonitor configuration has been refactored to provide a simpler interface:
- You no longer need to specify a monitor class in your
cromwell.confas this will now be inherited from thereference.confvalue. - You can now opt-in and opt-out of any combination of status monitors.
- The PAPI backends to monitor can now be listed in a single field.
You are no longer tied to the previous preset combinations of health checks. However if you just want to carry forward the exact same set of health checks, you can use one of the following standard recipes:
If you're currently using the (default) NoopHealthMonitorActor, no action is required.
If you're currently using the StandardHealthMonitorServiceActor, replace this stanza:
services {
HealthMonitor {
class = "cromwell.services.healthmonitor.impl.standard.StandardHealthMonitorServiceActor"
}
}
With this one:
services {
HealthMonitor {
config {
check-dockerhub: true
check-engine-database: true
}
}
}
Replace this stanza:
services {
HealthMonitor {
class = "cromwell.services.healthmonitor.impl.workbench.WorkbenchHealthMonitorServiceActor"
config {
papi-backend-name = PAPIv1
papi-v2-backend-name = PAPIv2
google-auth-name = service-account
gcs-bucket-to-check = "cromwell-ping-me-bucket"
}
}
}
With this one:
services {
HealthMonitor {
config {
check-dockerhub: true
check-engine-database: true
check-gcs: true
check-papi-backends: [PAPIv1, PAPIv2]
google-auth-name = service-account
gcs-bucket-to-check = "cromwell-ping-me-bucket"
}
}
}
A new workflow option is added. If the final_workflow_outputs_dir is set
use_relative_output_paths can be used. When set to true this will copy
all the outputs relative to their execution directory.
my_final_workflow_outputs_dir/~~MyWorkflow/af76876d8-6e8768fa/call-MyTask/execution/~~output_of_interest.
More information can be found in the workflow options documentation.
For example, the statement String s = "\"" is now supported, whereas previously it produced a syntax error.
Cromwell's WDL 1.0 implementation now allows empty call blocks, e.g. call task_with_no_inputs {}. This brings 1.0 in line with draft-2, which has always supported this syntax.
Fixed a bug that caused an error like Custom type was referred to but not found to be issued when using an imported type as a SchemaDefRequirement in packed CWL.
When set, the configuration value of system.cromwell_id will now have a random suffix appended, unless the
configuration key system.cromwell_id_random_suffix is set to false.
The generated id also appears more places in the logs, including when picking up workflows from the database and during shutdown.
Fixed an issue that caused the write_map() function in Cromwell's WDL 1.0 implementation to produce output in the wrong format. Specifically, the output's rows and columns were swapped. WDL draft-2 was not affected.
Incorrect write_map() output in Cromwell 38 and earlier:
key1 key2 key3
value1 value2 value3
Corrected write_map() output in Cromwell 39 and later:
key1 value1
key2 value2
key3 value3
The ConfigBackendLifecycleActorFactory path variables script, out and err are now consistent when running with
and without docker. Similarly, when killing a docker task the kill-docker configuration key is now used instead of
kill. For more information see the online documentation.
Previous versions of Cromwell defaulted to using a health monitor service that checked Docker Hub and engine database status.
Neither check was useful if the status endpoint was never consulted as is likely the case in most deployments. Cromwell 38
now defaults to a NoopHealthMonitorServiceActor which does nothing. The previous health service implementation is still
available as StandardHealthMonitorServiceActor.
- Fixed an issue that could cause Cromwell to consume disk space unnecessarily when using zipped dependencies
- When returning errors as json the
Content-Typeheader is set toapplication/json.
- Adds support for retrieving docker digests of asia.gcr.io images
- Adds configuration settings for docker digest lookups. See the
dockersection of thereference.conffor more information - Attempt to automatically adjust the boot disk size on the Google Cloud Backend (version 2) if the size of the image is greater than the default disk size or the required disk size in the runtime attributes. Only works for registries that support the version 2 of the manifest schema (https://docs.docker.com/registry/spec/manifest-v2-2/) At this date (12/09/18) this includes GCR and Dockerhub.
Call caching hashes with this new strategy are based on the path and the last modified time of the file.
For multi-instance Cromwell deployments sharing a single database, earlier versions of Cromwell required abort requests to be sent specifically to the instance that was running the targeted workflow. Cromwell 37 now allows abort commands to be sent to any Cromwell instance in the shared-database deployment. Configuration details here.
The Google Pipelines API (PAPI) version 1 and 2 backends now offer the option of call cache blacklisting on a per-bucket basis. More info here.
- All memory units in WDL are now treated as base-2.
For instance
1 KB == 1 KiB == 1024 Bytes.
Previous versions of Cromwell incorporated the name of the backend on which a call was run into the call cache hashes generated for that call.
Unfortunately this made it impossible to change the name of a backend without losing all previously run calls as potential cache hits.
Cromwell 37 introduces the name-for-call-caching-purposes backend configuration option as a means of decoupling the backend name from the
value used for the backend name for call caching purposes.
Support InputResourceRequirement hint
In cases where its not obvious why jobs are queued in Cromwell, you can enable logging for the Job Execution Token Dispenser, using
the system.hog-safety.token-log-interval-seconds configuration value.
The default, 0, means that no logging will occur.
- The HTTP filesystem is now enabled for engine use by default. To continue without an HTTP filesystem, you can add the following content into the appropriate stanza of your configuration file:
engine {
filesystems {
http {
enabled: false
}
}
}
- When the value
exit-code-timeout-secondsis set,check-alivecommand is now only called once every timeout interval instead of each poll.
This new endpoint brings the functionality of Womtool to the world of web services. Submit workflows for validation and receive a JSON description in response.
The endpoint is still undergoing heavy development and should not be used in production. The final version will ship in a future release of Cromwell; watch this space.
- Fixed a regression in Cromwell 36 that could cause operations on empty arrays to fail with a spurious type error (closes #4318)
On Hold workflows may now be aborted.
The AWS and TES backends can now handle calls that generate longer command lines. Like the other backends, commands scripts are first written to a file, the file is downloaded to the execution host, and then the localized script is run.
Also fixed are AWS command {} blocks that use | at the start of a line. For example:
command {
echo hello world \
| cat
}
The value exit-code-timeout-seconds can now set in a backend configuration.
Details here
Coalesce metadata requests to eliminate expensive and redundant queries and metadata construction.
Eliminate superfluous logging and metadata publishing in the shared filesystem backend on poll intervals where there was not a state change.
Previously US-EAST-1 was hardcoded in places.
Cromwell now allows for a user to submit the URL pointing to workflow file to run a workflow. More details on how to use it in:
- Added an opt-in namespace cache for the WDL Draft 2 language factory. Please see the Cromwell example configuration for details. NOTE: if upgrading from a hotfix version of Cromwell that relied upon this cache, the cache is now opt-in and must be turned on explicitly in config.
- To maintain conformance with the OpenWDL spec, Cromwell drops support for the
version draft-3identifier in this release. In the rare case where end users may have been usingversion draft-3,version 1.0is a drop-in replacement with no effect on functionality.
http and https workflow inputs are now supported for shared filesystem and Google Pipelines API (PAPI) version 2
backends. Configuration details are described here.
More efficient cache hit copying in multi-user environments is now supported through the call_cache_hit_path_prefixes workflow option.
Details here
Cromwell now offers the ability to cache file hashes on a root workflow level basis, details here.
The value dockerRoot can now be set in a backend configuration.
This will set the execution folder in the container (default: /cromwell-executions).
- The
releaseHoldendpoint will now return404 Not Foundfor an unrecognized workflow ID and400 Bad Requestfor a malformed or invalid workflow ID.
- Fixed a bug that allowed values to be "auto-boxed" into a single-element
Arrayof that type, which is not allowed in the WDL spec (Closes #3478).
- Restored standard output and error streaming for jobs.
- Fixes a bug which stopped
includeSubworkflow=falsefrom paging correctly and subworkflows from being discounted correctly fromtotalResultsCount. - Query results will now be returned in reverse chronological order, with the most-recently submitted workflows returned first.
Access of Google Cloud Storage buckets with Requester Pays enabled is now supported. Please read the relevant documentation for information on how to enable it and the consequences.
Support for private Docker Hub images is now included in the Google Pipelines API v2 backend. PAPI v2 private Docker support is equivalent to that in PAPI v1 but the configuration differs, please see Docker configuration for more details.
Updated the MySQL connector client from 5.1.42 to 5.1.46 which adds support for connecting to MySQL 8.0. See the
documentation on Changes in MySQL Connector/J for
more information.
This gives the ability to filter out workflows based on labels. Two new parameters called excludeLabelAnd and excludeLabelOr can be used for this purpose.
More details on how to use them can be found here.
Cromwell now supports excluding subworkflows from workflow query results using the includeSubworkflows parameter. By default they are included in the results.
More information can be found at REST API.
Cromwell now supports querying workflows by submission time. This will help find workflows that are submitted but not started yet (i.e. workflows which are in On Hold state). More information can be found here.
Submission time of a workflow is now included in WorkflowQueryResult, which is part of the response for workflow query.
Cromwell now allows tasks in WDL 1.0 can now specify an optimization in their parameter_meta that some File inputs do not need to be localized for the task to run successfully.
Full details are available in the documentation page for this optimization.
Workflows which are in 'On Hold' state can now be fetched using the query endpoint.
Initial support for Google Pipelines API version 2. Expect feature parity except for private dockerhub images which are not supported at the moment, but will be in the near future. Additionally, the "refresh token" authentication mode is NOT supported on PAPI V2.
In addition, the following changes are to be expected:
- Error messages for failed jobs might differ from V1
- The Pipelines API log file content might differ from V1
Important (If you're running Cromwell with a Google backend, read this):
The actor-factory value for the google backend (cromwell.backend.impl.jes.JesBackendLifecycleActorFactory) is being deprecated.
Please update your configuration accordingly.
| PAPI Version | actor-factory |
|---|---|
| V1 | cromwell.backend.google.pipelines.v1alpha2.PipelinesApiLifecycleActorFactory |
| V2 | cromwell.backend.google.pipelines.v2alpha1.PipelinesApiLifecycleActorFactory |
If you don't update the actor-factory value, you'll get a deprecation warning in the logs, and Cromwell will default back to PAPI V1
Cromwell now supports retrying failed tasks up to a specified count by declaring a value for the maxRetries key through the WDL runtime attributes.
- Cromwell has removed most of the formatting restrictions from custom labels. Please check the README for more detailed documentation.
- Custom labels won't be submitted to Google backend as they are now decoupled from Google's default labels.
- Cromwell now publishes the labels as soon as the workflow is submitted (whether started or on hold). If the labels are invalid, the workflow will not be submitted and request will fail.
From version 32 onwards we will no longer be publishing build artifacts compatible with Scala 2.11.
- If you don't import the classes into your own scala project then this should have no impact on you.
- If you are importing the classes into your own scala project, make sure you are using Scala 2.12.
Cromwell can now validate that your inputs files do not supply inputs with no impact on the workflow. Strict validation will be disabled by default in WDL draft 2 and CWL but enabled in WDL draft 3. See the 'Language Factory Config' below for details.
All language factories can now be configured on a per-language-version basis. All languages and versions will support the following options:
enabled: Defaults totrue. Set tofalseto disallow workflows of this language and version.strict-validation: Defaults totruefor WDL draft 3 andfalsefor WDL draft 2 and CWL. Specifies whether workflows fail if the inputs JSON (or YAML) file contains values which the workflow did not ask for (and will therefore have no effect). Additional strict checks may be added in the future.
- More accurately returns 503 instead of 500 when Cromwell can not respond in a timely manner
- Cromwell now allows a user to submit a workflow but in a state where it will not automatically be picked up for execution. This new state is called 'On Hold'. To do this you need to set the parameter workflowOnHold to true while submitting the workflow.
- API end point 'releaseHold' will allow the user to send a signal to Cromwell to allow a workflow to be startable, at which point it will be picked up by normal execution schemes.
The PAPI backend now supports specifying GPU through WDL runtime attributes:
runtime {
gpuType: "nvidia-tesla-k80"
gpuCount: 2
zones: ["us-central1-c"]
}The two types of GPU supported are nvidia-tesla-k80 and nvidia-tesla-p100
Important: Before adding a GPU, make sure it is available in the zone the job is running in: https://cloud.google.com/compute/docs/gpus/
Cromwell now allows for system-wide or per-backend job shell configuration for running user commands rather than always
using the default /bin/bash. To set the job shell on a system-wide basis use the configuration key system.job-shell or on a
per-backend basis with <config-key-for-backend>.job-shell. For example:
# system-wide setting, all backends get this
-Dsystem.job-shell=/bin/sh
# override for just the Local backend
-Dbackend.providers.Local.config.job-shell=/bin/sh
For the Config backend the value of the job shell will be available in the ${job_shell} variable. See Cromwell's reference.conf for an example
of how this is used for the default configuration of the Local backend.
The imports zip no longer unpacks a single (arbitrary) internal directory if it finds one (or more). Instead, import statements should now be made relative to the base of the import zip root.
Reverting to a prior custom label value now works.
"Retrieves the current labels for a workflow" will return the most recently summarized custom label value.
The above endpoint may still return the prior value for a short period of time after using "Updated labels for a workflow" until the background metadata summary process completes.
If you never used the REST API to revert a custom label back to a prior value you will not be affected. This only applies to workflows previously updated using "Updated labels for a workflow".
The database table storing custom labels will delete duplicate rows for any workflow label key. For efficiency purposes the values are not regenerated automatically from the potentially large metadata table.
In rare cases where one tried to revert to a prior custom label value you may continue to see different results
depending on the REST API used. After the database update
"Retrieves the current labels for a workflow"
will return the most-recent-unique value while
"Get workflow and call-level metadata for a specified workflow"
will return the up-to-date value. For example, if one previously updated a value from "value-1" > "value-2" >
"value-3" > "value-2" then the former REST API will return value-3 while the latter will return value-2.
Workflow metadata for jobs run on a Google Pipelines API backend will report the google_project specified via a
workflow options json.
-
Cromwell server
The Cromwell server source code is now located underserver/src.sbt assemblywill build the runnable Cromwell JAR inserver/target/scala-2.12/with a name likecromwell-<VERSION>.jar. -
Robustness
- The rate at which jobs are being started can now be controlled using the
system.job-rate-controlconfiguration stanza. - A load controller service has been added to allow Cromwell to self-monitor and adjust its load accordingly.
The load controller is currently a simple on/off switch controlling the job start rate. It gathers metrics from different parts of the system
to inform its decision to stop the creation of jobs.
You can find relevant configuration in the
services.LoadControllersection of thecromwell.examples.conffile, as well as in theload-controlsection inreference.conf. The load level of the monitored sub-systems are instrumented and can be found under thecromwell.loadstatsD path. - The statsD metrics have been re-shuffled a bit. If you had a dashboard you might find that you need to update it.
Changes include:
- Removed artificially inserted "count" and "timing" the path
- Added a
loadsection - Metrics were prefixed twice with
cromwell(cromwell.cromwell.my_metric), now they're only prefixed once - Added
processedandqueuemetrics under various metrics monitoring the throughput and amount of queued work respectively - Added a memory metric representing an estimation of the free memory Cromwell thinks it has left
- The rate at which jobs are being started can now be controlled using the
-
Added a configuration option under
docker.hash-lookup.enabledto disable docker hash lookup. Disabling it will also disable call caching for jobs with floating docker tags. -
API
- Updated the
/queryresponse to include the total number of query results returned. See here for more information.
- Updated the
- A set of bug fixes following the migration of Cromwell to WOM (the Workflow Object Model) in version 30.
- The
customLabelsform field for workflow submission has been renamed tolabels.
-
New Cromwell documentation
Our documentation has moved from our README to a new website: Cromwell Documentation. There are new Tutorials and much of the documentation has been re-written. The source files are in the /docs directory. -
API
- Cromwell now supports input files in the yaml format (JSON format is still supported).
- Added a GET version for the
labelsendpoint which will return current labels for a workflow.
-
Database
You have the option of storing the metadata in a separate SQL database than the database containing the internal engine data. When switching connection information for an existing database containing historical data, the tables should be manually replicated from one database instance to another using the tools appropriate for your specific database types. Cromwell will not move any existing data automatically. This feature should be considered experimental and likely to change in the future. See the Database Documentation or thedatabasesection in cromwell.examples.conf for more information. -
StatsD
Added initial support for StatsD instrumentation. See the Instrumentation Documentation for details on how to use it. -
User Service Account auth mode for Google
Added a new authentication mode for Google Cloud Platform which will allow a user to supply the JSON key file in their workflow options to allow for per-workflow authentication via service account. This is analogous to the previously existing refresh token authentication scheme. As with the refresh token scheme it is encouraged that the user_service_account_json workflow option field is added to the encrypted-fields list in the configuration. -
Bugfixes
Abort of Dockerized tasks on the Local backend should now work as expected. Cromwell usesdocker killto kill the Docker container.
-
Command line
In preparation for supporting CWL scripts (yes, you read that right!), we have extensively revised the Command Line in Cromwell 29. For more details about the usage changes please see the README. And stay tuned to the WDL/Cromwell blog over the next couple of months for more news about CWL. -
Request timeouts
Cromwell now returns more specific503 Service Unavailableerror codes on request timeouts, rather than the more generic500 Internal Server Error. The response for a request timeout will now be plain text, rather than a JSON format. -
Metadata endpoint
The response from the metadata endpoint can be quite large depending on your workflow. You can now opt-in to have Cromwell gzip your metadata file, in order to reduce file size, by sending theAccept-Encoding: gzipheader. The default behavior now does not gzip encode responses. -
Engine endpoints
Previously the engine endpoints were available under/api/enginebut now the endpoints are under/engineso they don't require authentication. Workflow endpoints are still available under/api/workflows. We also deprecated the settingapi.routeUnwrappedas a part of this internal consistency effort. -
Call caching diff
We updated the response format of the callcaching/diff endpoint.
-
Cromwell server
When running in server mode, Cromwell now attempts to gracefully shutdown after receiving aSIGINT(Ctrl-C) orSIGTERM(kill) signal. This means that Cromwell waits for all pending database writes before exiting, as long as you includeapplication.confat the top of your config file. You can find detailed information about how to configure this feature in the Cromwell Wiki. -
Concurrent jobs
You can now limit the number of concurrent jobs for any backend. Previously this was only possible in some backend implementations. Please see the README for details.
-
Optional WDL variables
Empty optional WDL values are now rendered as thenullJSON value instead of the JSON string"null"in the metadata and output endpoints. You do not need to migrate previous workflows. Workflows run on Cromwell 28 and prior will still render empty values as"null". -
Empty WDL variables
Cromwell now acceptsnullJSON values in the input file and coerces them as an empty WDL value. WDL variables must be declared optional in order to be supplied with anullJSON value.
input.json
{
"null_input_values.maybeString": null,
"null_input_values.arrayOfMaybeInts": [1, 2, null, 4]
}workflow.wdl
workflow null_input_values {
String? maybeString
Array[Int?] arrayOfMaybeInts
}
The following WDL functions now add a newline after the final line of output (the previous behavior of not adding this newline was inadvertent):
write_lineswrite_mapwrite_objectwrite_objectswrite_tsv
For example:
task writer {
Array[String] a = ["foo", "bar"]
command {
# used to output: "foo\nbar"
# now outputs: "foo\nbar\n"
cat write_lines(a)
}
}
A workflow utilizing the WorkflowFailureMode Workflow Option ContinueWhilePossible will now successfully reach a terminal state once all runnable jobs have completed.
When FailOnStderr is set to false, Cromwell no longer checks for the existence of a stderr file for that task.
Enables the floor, ceil and round functions in WDL to convert floating point numbers to integers.
For example we can now use the size of an input file to influence the amount of memory the task is given. In the example below a 500MB input file will result in a request for a VM with 2GB of memory:
task foo {
File in_file
command { ... }
runtime {
docker: "..."
memory: ceil(size(in_file)) * 4
}
}
- Hash values calculated by Cromwell for a call when call caching is enabled are now published to the metadata.
It is published even if the call failed. However if the call is attempted multiple times (because it has been preempted for example),
since hash values are strictly identical for all attempts, they will only be published in the last attempt section of the metadata for this call.
If the hashes fail to be calculated, the reason is indicated in a
hashFailuresfield in thecallCachingsection of the call metadata. Important: Hashes are not retroactively published to the metadata. Which means only workflows run on Cromwell 28+ will have hashes in their metadata.
See the README for an example metadata response.
- New endpoint returning the hash differential for 2 calls.
GET /api/workflows/:version/callcaching/diff
See the README for more details.
- The workflow submission parameters
wdlSourceandwdlDependencieshave been deprecated in favor ofworkflowSourceandworkflowDependenciesrespectively. The older names are still supported in Cromwell 28 with deprecation warnings but will be removed in a future version of Cromwell.
- A new
/labelsendpoint has been added to update labels for an existing workflow. See the README for more information. - Label formatting requirements have been updated, please check the README for more detailed documentation.
The JES backend now supports a filesystems.gcs.caching.duplication-strategy configuration entry.
It can be set to specify the desired behavior of Cromwell regarding call outputs when a call finds a hit in the cache.
The default value is copy which will copy all output files to the new call directory.
A second value is allowed, reference, that will instead point to the original output files, without copying them.
filesystems {
gcs {
auth = "application-default"
caching {
duplication-strategy = "reference"
}
}
}A placeholder file will be placed in the execution folder of the cached call to explain the absence of output files and point to the location of the original ones.
Metadata write batching works the same as in previous versions of Cromwell, but the default batch size has been changed from 1 to 200. It's possible that 200 is too high in some environments, but 200 is more likely to be an appropriate value than the previous default.
- Call Caching has been improved in this version of Cromwell, specifically the time needed to determine whether or not a job can be cached
has drastically decreased. To achieve that the database schema has been modified and a migration is required in order to preserve the pre-existing cached jobs.
This migration is relatively fast compared to previous migrations. To get an idea of the time needed, look at the size of your
CALL_CACHING_HASH_ENTRYtable. As a benchmark, it takes 1 minute for a table with 6 million rows. The migration will only be executed on MySQL. Other databases will lose their previous cached jobs. In order to run properly on MySQL, the following flag needs to be adjusted: https://dev.mysql.com/doc/refman/5.5/en/server-system-variables.html#sysvar_group_concat_max_len The following query will give you a minimum to set the group_concat_max_len value to:
SELECT MAX(aggregated) as group_concat_max_len FROM
(
SELECT cche.CALL_CACHING_ENTRY_ID, SUM(LENGTH(CONCAT(cche.HASH_KEY, cche.HASH_VALUE))) AS aggregated
FROM CALL_CACHING_HASH_ENTRY cche
GROUP BY cche.CALL_CACHING_ENTRY_ID
) aggregationHere is the SQL command to run to set the group_concat_max_len flag to the proper value:
SET GLOBAL group_concat_max_len = valueWhere value is replaced with the value you want to set it to.
Note that the migration will fail if the flag is not set properly.
- The update to Slick 3.2 requires a database stanza to
switch from using
drivertoprofile.
database {
#driver = "slick.driver.MySQLDriver$" #old
profile = "slick.jdbc.MySQLProfile$" #new
db {
driver = "com.mysql.cj.jdbc.Driver"
url = "jdbc:mysql://host/cromwell?rewriteBatchedStatements=true"
user = "user"
password = "pass"
connectionTimeout = 5000
}
}Cromwell now supports call caching with floating Docker tags (e.g. docker: "ubuntu:latest"). Note it is still considered
a best practice to specify Docker images as hashes where possible, especially for production usages.
Within a single workflow Cromwell will attempt to resolve all floating tags to the same Docker hash, even if Cromwell is restarted
during the execution of a workflow. In call metadata the docker runtime attribute is now the same as the
value that actually appeared in the WDL:
"runtimeAttributes": {
"docker": "ubuntu:latest",
"failOnStderr": "false",
"continueOnReturnCode": "0"
}
Previous versions of Cromwell rewrote the docker value to the hash of the Docker image.
There is a new call-level metadata value dockerImageUsed which captures the hash of the Docker image actually used to
run the call:
"dockerImageUsed": "library/ubuntu@sha256:382452f82a8bbd34443b2c727650af46aced0f94a44463c62a9848133ecb1aa8"
- The Docker section of the configuration has been slightly reworked
An option to specify how a Docker hash should be looked up has been added. Two methods are available.
"local" will try to look for the image on the machine where cromwell is running. If it can't be found, Cromwell will try to
pullthe image and use the hash from the retrieved image. "remote" will try to look up the image hash directly on the remote repository where the image is located (Docker Hub and GCR are supported) Note that the "local" option will require docker to be installed on the machine running cromwell, in order for it to call the docker CLI. - Adds hash lookup support for public quay.io images.
- Added support for the new WDL
basenamefunction. Allows WDL authors to get just the file name from a File (i.e. removing the directory path) - Allows coercion of
Mapobjects intoArrays ofPairs. This also allows WDL authors to directly scatter over WDLMaps.
- Adds support for JSON file format for google service account credentials. As of Cromwell 27, PEM credentials for PAPI are deprecated and support might be removed in a future version.
google {
application-name = "cromwell"
auths = [
{
name = "service-account"
scheme = "service_account"
json-file = "/path/to/file.json"
}
]
}
- The
/queryendpoint now supports querying bylabel. See the README for more information. - The
read_Xstandard library functions limit accepted filesizes. These differ by type, e.g. read_bool has a smaller limit than read_string. See reference.conf for default settings.
- Failure metadata for calls and workflows was being displayed inconsistently, with different formats depending on the originating Cromwell version. Failures will now always present as an array of JSON objects each representing a failure. Each failure will have a message and a causedBy field. The causedBy field will be an array of similar failure objects. An example is given below:
failures: [{
message: "failure1",
causedBy: [{
message: "cause1",
causedBy: []
}, {
message: "cause2",
causedBy: []
}]
}, {
message: "failure2",
causedBy: []
}]
- Upgrading to Cromwell 26 will take additional time due to the migration of failure metadata. Cromwell will automatically run a database query during the upgrade which appears to be roughly linear to the number of rows in the METADATA_ENTRY table. You can estimate upgrade time using the following equation:
time to migrate (in seconds) ~= (rows in METADATA_ENTRY) / 65000Note that due to differences in hardware and database speed, this is only a rough estimate.
- Added a configuration option under
system.ioto throttle the number of I/O queries that Cromwell makes, as well as configure retry parameters. This is mostly useful for the JES backend and should be updated to match the GCS quota available for the project.
system.io {
# Global Throttling - This is mostly useful for GCS and can be adjusted to match
# the quota availble on the GCS API
number-of-requests = 100000
per = 100 seconds
# Number of times an I/O operation should be attempted before giving up and failing it.
number-of-attempts = 5
}
- A special thank you to @adamstruck, @antonkulaga and @delocalizer for their contributions to Cromwell.
- Metadata keys for call caching are changed. All call caching keys are now in a
callCachingstanza.Call cache read resulthas moved here and is nowresult. TheallowResultReuseandeffectiveCallCachingModehave moved here. Thehitboolean is a simple indication of whether or not it was a hit, with no additional information. An example using the new format is:
"callCaching": {
"hit": false,
"effectiveCallCachingMode": "ReadAndWriteCache",
"result": "Cache Miss",
"allowResultReuse": true
}
- Added a field
insert-batch-sizeto thedatabasestanza which defines how many values from a batch insert will be processed at a time. This value defaults to 2000. - Moved the config value
services.MetadataService.metadata-summary-refresh-intervaltoservices.MetadataService.config.metadata-summary-refresh-interval - Added ability to override the default zone(s) used by JES via the config structure by setting
genomics.default-zonesin the JES configuration - The cromwell server TCP binding timeout is now configurable via the config key
webservice.binding-timeout, defaulted to the previous value5s(five seconds) via the reference.conf. - For MySQL users, a massive scalability improvement via batched DB writing of internal metadata events. Note that one must add
rewriteBatchedStatements=trueto their JDBC URL in their config in order to take advantage of this
- Cromwell's WDL parser now recognizes empty array literals correctly, e.g.
Array[String] emptyArray = []. - Cromwell now applies default labels automatically to JES pipeline runs.
- Added support for new WDL functions:
length: (Array[X]) => Integer- report the length of the specified arrayprefix: (String, Array[X]) => Array[String]- generate an array consisting of each element of the input array prefixed by a specifiedString. The input array can have elements of any primitive type, the return array will always have typeArray[String].defined: (Any) => Boolean- Will return false if the provided value is an optional that is not defined. Returns true in all other cases.
- Cromwell's Config (Shared Filesystem) backend now supports invocation of commands which run in a Docker image as a non-root user.
The non-root user could either be the default user for a given Docker image (e.g. specified in a Dockerfile via a
USERdirective), or the Config backend could pass an optional"-u username"as part of thesubmit-dockercommand. - In some cases the SFS backend, used for Local, SGE, etc., coerced
WdlFiletoWdlStringby using.toUri. This resulted in strings prepended withfile:///path/to/file. Now absolute file paths will not contain the uri scheme. - Launch jobs on servers that support the GA4GH Task Execution Schema using the TES backend.
- Call caching: Cromwell will no longer try to use the cache for WDL tasks that contain a floating docker tag. Call caching will still behave the same for tasks having a docker image with a specific hash. See https://github.com/broadinstitute/cromwell#call-caching-docker-tags for more details.
- Added docker hash lookup. Cromwell will try to lookup the hash for a docker image with a floating tag, and use that hash when executing the job. This will be reflected in the metadata where the docker runtime attribute will contains the hash that was used. If Cromwell is unable to lookup the docker hash, the job will be run with the original user defined floating tag. Cromwell is currently able to lookup public and private docker hashes for images on Docker Hub and Google Container Engine for job running on the JES backend. For other backends, cromwell is able to lookup public docker hashes for Docker Hub and Google Container Engine. See https://github.com/broadinstitute/cromwell#call-caching-docker-tags for more details.
- Added CUSTOM_LABELS as a field of WORKFLOW_STORE_ENTRY, to store workflow store entries.
- When emitting workflow outputs to the Cromwell log only the first 1000 characters per output will be printed
- Added support for conditional (
if) statements. - Globs for Shared File System (SFS) backends, such as local or SGE, now use bash globbing instead of Java globbing, consistent with the JES backend.
- The
metaandparameter_metablocks are now valid withinworkflowblocks, not justtask - The JES backend configuration now has an option
genomics-api-queries-per-100-secondsto help tune the rate of batch polling against the JES servers. Users with quotas larger than default should make sure to set this value. - Added an option
call-caching.invalidate-bad-cache-results(default:true). If true, Cromwell will invalidate cached results which have failed to copy as part of a cache hit. - Timing diagrams and metadata now receive more fine grained workflow states between submission and Running.
- Support for the Pair WDL type (e.g.
Pair[Int, File] floo = (3, "gs://blar/blaz/qlux.txt")) - Added support for new WDL functions:
zip: (Array[X], Array[Y]) => Array[Pair[X, Y]]- align items in the two arrays by index and return them as WDL pairscross: (Array[X], Array[Y]) => Array[Pair[X, Y]]- create every possible pair from the two input arrays and return them all as WDL pairstranspose: (Array[Array[X]]) => Array[Array[X]]compute the matrix transpose for a 2D array. Assumes each inner array has the same length.
- By default,
system.abort-jobs-on-terminateis false when runningjava -jar cromwell.jar server, and true when runningjava -jar cromwell.jar run <wdl> <inputs>. - Enable WDL imports when running in Single Workflow Runner Mode.
- Both batch and non-batch REST workflow submissions now require a multipart/form-data encoded body.
- Support for sub workflows (see Annex A)
- Enable WDL imports when running in Single Workflow Runner Mode as well as Server Mode
- Support for WDL imports through an additional imports.zip parameter
- Support for sub workflows
- Corrected file globbing in JES to correctly report all generated files. Additionally, file globbing in JES now uses bash-style glob syntax instead of python style glob syntax
- Support declarations as graph nodes
- Added the ability to override the default service account that the compute VM is started with via the configuration option
JES.config.genomics.compute-service-accountor through the workflow options parametergoogle_compute_service_account. More details can be found in the README.md - Fix bugs related to the behavior of Cromwell in Single Workflow Runner Mode. Cromwell will now exit once a workflow completes in Single Workflow Runner Mode. Additionally, when restarting Cromwell in Single Workflow Runner Mode, Cromwell will no longer restart incomplete workflows from a previous session.
The WDL specification has changed regarding workflow outputs to accommodate sub workflows. This change is backward compatible in terms of runnable WDLs (WDL files using the deprecated workflow outputs syntax will still run the same). The only visible change lies in the metadata (as well as the console output in single workflow mode, when workflow outputs are printed out at the end of a successful workflow).
TL;DR Unless you are parsing or manipulating the "key" by which workflow outputs are referenced in the metadata (and/or the console output for single workflow mode), you can skip the following explanation.
Metadata Response
{
...
outputs {
"task_output_1": "hello",
"task_output_2": "world"
^
If you don't manipulate this part of the metadata, then skip this section
}
}
In order to maintain backward compatibility, workflow outputs expressed with the deprecated syntax are "expanded" to the new syntax. Here is an example:
task t {
command {
#do something
}
output {
String out1 = "hello"
String out2 = "world"
}
}
workflow old_syntax {
call t
output {
t.*
}
}
workflow new_syntax {
call t
output {
String wf_out1 = t.out1
String wf_out2 = t.out2
}
}
The new syntax allows for type checking of the outputs as well as expressions. It also allows for explicitly naming to the outputs. The old syntax doesn't give the ability to name workflow outputs. For consistency reasons, Cromwell will generate a "new syntax" workflow output for each task output, and name them. Their name will be generated using their FQN, which would give
output {
String w.t.out1 = t.out1
String w.t.out2 = t.out2
}
However as the FQN separator is ., the name itself cannot contain any ..
For that reason, . are replaced with _ :
Old syntax expanded to new syntax
output {
String w_t_out1 = t.out1
String w_t_out2 = t.out2
}
The consequence is that the workflow outputs section of the metadata for old_syntax would previously look like
outputs {
"w.t.out1": "hello",
"w.t.out2": "hello"
}
but it will now look like
outputs {
"w_t_out1": "hello",
"w_t_out2": "hello"
}
The same applies for the console output of a workflow run in single workflow mode.
- Improved retries for Call Caching and general bug fixes.
- Users will experience better scalability of status polling for Google JES.
- Now there are configurable caching strategies for a SharedFileSystem backend (i.e. Local, SFS) in the backend's stanza: See below for detailed descriptions of each configurable key.
backend {
...
providers {
SFS_BackendName {
actor-factory = ...
config {
...
filesystems {
local {
localization: [
...
]
caching {
duplication-strategy: [
"hard-link", "soft-link", "copy"
]
# Possible values: file, path
# "file" will compute an md5 hash of the file content.
# "path" will compute an md5 hash of the file path. This strategy will only be effective if the duplication-strategy (above) is set to "soft-link",
# in order to allow for the original file path to be hashed.
hashing-strategy: "file"
# When true, will check if a sibling file with the same name and the .md5 extension exists, and if it does, use the content of this file as a hash.
# If false or the md5 does not exist, will proceed with the above-defined hashing strategy.
check-sibling-md5: false
}
-
Multiple Input JSON files can now be submitted in server mode through the existing submission endpoint: /api/workflows/:version. This endpoint accepts a POST request with a multipart/form-data encoded body. You can now include multiple keys for workflow inputs.
Each key below can contain an optional JSON file of the workflow inputs. A skeleton file can be generated from wdltool using the "inputs" subcommand. NOTE: In case of key conflicts between multiple JSON files, higher values of x in workflowInputs_x override lower values. For example, an input specified in workflowInputs_3 will override an input with the same name that was given in workflowInputs or workflowInputs_2. Similarly, an input specified in workflowInputs_5 will override an input with the same name in any other input file. workflowInputs workflowInputs_2 workflowInputs_3 workflowInputs_4 workflowInputs_5 -
You can now limit the number of concurrent jobs for a backend by specifying the following option in the backend's config stanza:
backend {
...
providers {
BackendName {
actor-factory = ...
config {
concurrent-job-limit = 5
-
Warning: Significant database updates when you switch from version 0.19 to 0.21 of Cromwell. There may be a long wait period for the migration to finish for large databases. Please refer to MIGRATION.md for more details.
-
There are significant architectural changes related to increases in performance and scaling.
-
The biggest user-facing changes from 0.19 to 0.21 are related to the application.conf file, which has been restructured significantly. The configuration for backends now is all contained within a
backendstanza, which specifies 1 stanza per name per backend and a default backend, as follows:
backend {
default=Local
providers {
Local {
actor-factory = "cromwell.backend.impl.sfs.config.ConfigBackendLifecycleActorFactory"
config {
... backend specific config ...
}
}
JES {
actor-factory = "cromwell.backend.impl.jes.JesBackendLifecycleActorFactory"
config {
... backend specific config ...
}
}
SGE {
actor-factory = "cromwell.backend.impl.sfs.config.ConfigBackendLifecycleActorFactory"
config {
... backend specific config ...
}r
}
}
}
-
A new
/statsendpoint has been added to get workflow and job count for a Cromwell running in server mode. -
Renamed Workflow Options: “workflow_log_dir” -> “final_workflow_log_dir” “call_logs_dir” -> “final_call_logs_dir” “outputs_path” -> “final_workflow_outputs_dir” “defaultRuntimeOptions” -> “default_runtime_attributes”
-
Timing diagrams endpoint has been updated to include additional state information about jobs.
-
Add support for Google Private IPs through
noAddressruntime attribute. If set to true, the VM will NOT be provided with a public IP address. Important: Your project must be whitelisted in "Google Access for Private IPs Early Access Program". If it's not whitelisted and you set this attribute to true, the task will hang. Defaults tofalse. e.g:
task {
command {
echo "I'm private !"
}
runtime {
docker: "ubuntu:latest"
noAddress: true
}
}
-
The Local and the SGE backend have been merged into a generic Shared File System (SFS) backend. This updated backend can be configured to work with various other command line dispatchers such as LSF. See the README for more info.
-
On the JES and SFS backends, task
commandblocks are now always passed absolute paths for inputFiles. -
On the SFS backends, the call directory now contains two sub-directories:
inputscontains all the input files that have been localized for this task (see next below for more details)executioncontains all other files (script, logs, rc, potential outputs etc...)
-
Override the default database configuration by setting the keys
database.driver,database.db.driver,database.db.url, etc. -
Override the default database configuration by setting the keys
database.driver,database.db.driver,database.db.url, etc.
For example:
# use a mysql database
database {
driver = "slick.driver.MySQLDriver$"
db {
driver = "com.mysql.cj.jdbc.Driver"
url = "jdbc:mysql://host/cromwell"
user = "user"
password = "pass"
connectionTimeout = 5000
}
}
-
The default per-upload bytes size for GCS is now the minimum 256K instead of 64M. There is also an undocumented config key
google.upload-buffer-bytesthat allows adjusting this internal value. -
Updated Docker Hub hash retriever to parse json with custom media types.
-
Added a
/batchsubmit endpoint that accepts a single wdl with multiple input files. -
The
/queryendpoint now supports querying byid, and submitting parameters as a HTTP POST.