Unable to load partitions in Athena with case sensitivity ON

Question

I have data in S3 which is partitioned in YYYY/MM/DD/HH/ structure (not year=YYYY/month=MM/day=DD/hour=HH)

I set up a Glue crawler for this, which creates a table in Athena, but when I query the data in Athena it gives an error as one field has duplicate name (URL and url , which the SerDe converts to lowercase, causing a name conflict).

To fix this, I manually create another table (using the above table definition SHOW CREATE TABLE), adding 'case.insensitive'= FALSE to the SERDEPROPERTIES

WITH SERDEPROPERTIES ('paths'='deviceType,emailId,inactiveDuration,pageData,platform,timeStamp,totalTime,userId','case.insensitive'= FALSE) 

I changed the s3 directory structure to the hive-compatible naming year=/month=/day=/hour= and then created the table with 'case.insensitive'= FALSE, then ran the MSCK REPAIR TABLE command for the new table, which loads all the partitions. (Complete CREATE TABLE QUERY)

But upon querying, I can only find 1 data column (platform) and the partition columns, rest of all the columns are not parsed. But I've actually copied the Glue-generated CREATE TABLE query, with the case_insensitive=false condition.

How can I fix this?

Answer 1

I think you have multiple, separate issues: one with the crawler, and one with the serde, and one with duplicate keys:

Glue Crawler

If Glue Crawler delivered on what they promise they would be a fairly good solution for most situations and would save us from writing the same code over and over again. Unfortunately, if you stray outside of the (undocumented) use cases Glue Crawler was designed for, you often end up with various issues, from the strange to the completely broken (see for example this question, this question, this question, this question, this question, or this question).

I recommend that you skip Glue Crawler and instead write the table DDL by hand (you have a good template in what the crawler created, it just isn't good enough). Then you write a Lambda function (or shell script) that you run on a schedule to add new partitions.

Since your partitioning is only on time, this is a fairly simple script: it just needs to run every once in a while and add the partition for the next period.

It looks like your data is from Kinesis Data Firehose which produces a partitioned structure at hour granularity. Unless you have lots of data coming every hour I recommend you create a table that is only partitioned on date, and run the Lambda function or script once per day to add the next day's partition.

A benefit from not using Glue Crawler is that you don't have to have a one-to-one correspondence between path components and partition keys. You can have a single partition key that is typed as date, and add partitions like this: ALTER TABLE foo ADD PARTITION (dt = '2020-05-13') LOCATION 's3://some-bucket/data/2020/05/13/'. This is convenient because it's much easier to do range queries on a full date than when the components are separate.

If you really need hourly granularity you can either have two partition keys, one which is the date and one the hour, or just the one with the full timestamp, e.g. ALTER TABLE foo ADD PARTITION (ts = '2020-05-13 10:00:00') LOCATION 's3://some-bucket/data/2020/05/13/10/'. Then run the Lambda function or script every hour, adding the next hour's partition.

Having too a granular partitioning doesn't help with performance, and can instead hurt it (although the performance hit comes mostly from the small files and the directories).

SerDe config

As for the reason why you're only seeing the value of the platform column, it's because it's the only case where the column name and property have the same casing.

It's a bit surprising that the DDL you link to doesn't work, but I can confirm that it really doesn't. I tried creating a table from that DDL, but without the pagedata column (I also skipped the partitioning, but that shouldn't make a difference for the test), and indeed only the platform column had any value when I queried the table.

However, when I removed the case.insensitive serde property it worked as expected, which got me thinking that it might not work the way you think it does. I tried setting it to TRUE instead of FALSE, which made the table work as expected again. I think we can conclude from this that the Athena documentation is just wrong when it says "By default, Athena requires that all keys in your JSON dataset use lowercase". In fact, what happens is that Athena lower cases the column names, but it also lower cases the property names when reading the JSON.

With further experimentation it turned out the path property was redundant too. This is a table that worked for me:

CREATE EXTERNAL TABLE `json_case_test` (
  `devicetype` string, 
  `timestamp` string, 
  `totaltime` string, 
  `inactiveduration` int, 
  `emailid` string, 
  `userid` string, 
  `platform` string
)
ROW FORMAT SERDE 'org.openx.data.jsonserde.JsonSerDe' 
STORED AS INPUTFORMAT 'org.apache.hadoop.mapred.TextInputFormat' 
OUTPUTFORMAT 'org.apache.hadoop.hive.ql.io.HiveIgnoreKeyTextOutputFormat'
LOCATION 's3://some-bucket/data/'

I'd say that case.insensitive seems to cause more problems than it solves.

Duplicate keys

When I added the pagedata column (as struct<url:string>) and added "pageData":{"URL":"URL","url":"url"} to the data, I got the error:

HIVE_CURSOR_ERROR: Row is not a valid JSON Object - JSONException: Duplicate key "url"

And I got the error regardless of whether the pagedata column was involved in the query or not (e.g. SELECT userid FROM json_case_test also errored). I tried the case.insensitive serde property with both TRUE and FALSE, but it had no effect.

Next, I took a look at the source documentation for the serde, which first of all is worded much better, and secondly contains the key piece of information: that you also need to provide mappings for the columns when you turn off case insensitivity.

With the following serde properties I was able to get the duplicate key issue to go away:

WITH SERDEPROPERTIES (
  "case.insensitive" = "false",
  "mapping.pagedata" = "pageData",
  "mapping.pagedata.url" = "pagedata.url",
  "mapping.pagedata.url2"= "pagedata.URL"
)

You would have to provide mappings for all the columns except for platform, too.

---

Alternative: use JSON functions

You mentioned in a comment to this answer that the schema of the pageData property is not constant. This is another case where Glue Crawlers unfortunately don't really work. If you're unlucky you'll end up with a flapping schema that includes some properties some days (see for example this question).

What I realised when I saw your comment is that there is another solution to your problem: set up the table manually (as described above) and use string as the type for the pagedata column. Then you can use functions like JSON_EXTRACT_SCALAR to extract the properties you want during query time.

This solution trades increased complexity of the queries for way fewer headaches trying to keep up with an evolving schema.