spark | Firely and Incendi 's open source FHIR server

This is a workaround though and Glue Dev team is working on a fix, although there is no ETA.

Also this is still very buggy. You can not call .show() on a DynamicFrame for example, you need to call it on a DataFrame. Also all my jobs failed where I call data_frame.rdd.isEmpty(), don't ask me why.

Update 24.11.2021: I reached out to the Glue Dev Team and they told me that this is the intended way of fixing it. There is a workaround that can be done inside of the script though:

This usually happens, when various lib bring in different version of same lib. In this case it is Jackson. java.lang.NoSuchMethodError: com.fasterxml.jackson.dataformat.xml.XmlMapper.coercionConfigDefaults()Lcom/fasterxml/jackson/databind/cfg/MutableCoercionConfig; means: One lib probably require Jackson version, which has this method, but on class path is version, which does not yet have this funcion or got removed bcs was deprecated or renamed.

In case like this is good to print dependency tree and check version of Jackson required in libs. And if possible use newer versions of requid libs.

Solution: use libs, which use compatible versions of Jackson lib. No other shortcut possible.

• https://github.com/docker/compose/issues/3874

The thread points out that this is the expected behavior and is documented here: https://docs.docker.com/compose/extends/#understanding-multiple-compose-files

When you use multiple configuration files, you must make sure all paths in the files are relative to the base Compose file (the first Compose file specified with -f). This is required because override files need not be valid Compose files. Override files can contain small fragments of configuration. Tracking which fragment of a service is relative to which path is difficult and confusing, so to keep paths easier to understand, all paths must be defined relative to the base file.

There's a workaround within that discussion that works fairly well: https://github.com/docker/compose/issues/3874#issuecomment-470311052

The workaround is to use a ENV var that has a default:

• ${PROXY:-.}/haproxy/conf:/usr/local/etc/haproxy

Or in my case:

After reading the data frame, you can use withColumn() function to rename the date field.

Something like the following should work

When parsing raw files, you have a couple of options you can consider:

1. ❌ You can write your own parser to read bytes from files and convert them into data Spark can understand.
   • This is highly discouraged whenever possible due to the engineering time and unscalable architecture. It doesn't take advantage of distributed compute when you do this as you must bring the entire raw file to your parsing method before you can use it. This is not an effective use of your resources.
2. ⚠ You can use your own parser library not made for Spark, such as the XML Python library mentioned in the question
   • While this is less difficult to accomplish than writing your own parser, it still does not take advantage of distributed computation in Spark. It is easier to get something running, but it will eventually hit a limit of performance because it does not take advantage of low-level Spark functionality only exposed when writing a Spark library.
3. ✅ You can use a Spark-native raw file parser
   • This is the preferred option in all cases as it takes advantage of low-level Spark functionality and doesn't require you to write your own code. If a low-level Spark parser exists, you should use it.

In our case, we can use the Databricks parser to great effect.

In general, you should also avoid using the .udf method as it likely is being used instead of good functionality already available in the Spark API. UDFs are not as performant as native methods and should be used only when no other option is available.

A good example of UDFs covering up hidden problems would be string manipulations of column contents; while you technically can use a UDF to do things like splitting and trimming strings, these things already exist in the Spark API and will be orders of magnitude faster than your own code.

Our design is going to use the following:

1. Low-level Spark-optimized file parsing done via the Databricks XML Parser
2. Test-driven raw file parsing as explained here

First, we need to add the .jar to our spark_session available inside Transforms. Thanks to recent improvements, this argument, when configured, will allow you to use the .jar in both Preview/Test and at full build time. Previously, this would have required a full build but not so now.

We need to go to our transforms-python/build.gradle file and add 2 blocks of config:

1. Enable the pytest plugin
2. Enable the condaJars argument and declare the .jar dependency

My /transforms-python/build.gradle now looks like the following:

Just add to your Dockerfile this setting before running npm install: