Salesforce Batch Sink

The salesforce batch sink is responsible for using Salesforce API to upsert salesforce objects. The sink should handle large batches of data (~10 GB) and should handle all object types - contacts, campaigns, oppurtunities, leads, custom objects. Users should be able to upload subset of fields 

Use Cases 

• As a ETL developer, I would like to consolidate contacts data from multiple source systems and update in Salesforce

• As a ETL developer, I would like to aggregate results of a campaign from data warehouse into Salesforce

Assumptions

• Compound fields such as address and geo location will be represented as nested records in the source system

User Expectations 

• Users should be able configure the login URL
  • Hints should be provide in the default value 
• Users should be able configure the sensitive information such as client id, secret in secure store 
• Any error in the upload fails the pipeline
• Field level metadata should be captured by the sink 
  
  

User Configurations 

[1] Salesforce Bulk API will respond with result entry for inserting every single record. Either will "record **ID** was inserted" or with "error: **error_message**". During my testing it happened pretty often that insert was partially successful. For example:

• for records with some field empty it's SUCCESS, but for records where it's not, it says field is not insertable of wrong type etc.
• insertion of part of records was successful, while other failed due to "Storage limit exceeded".
• some field is required so records where it's empty will get an insertion error.

[2] defaults for "Maximum bytes per batch" and "Maximum records per batch" are taken from examples in Salesforce documenation.

[3] according to Bulk API Limitations batch will fail if it has more than 10 MB. So if user sets maximum bytes to something more than 10 millions we should tell the user that it does not make sense by failing the pipeline.

[4] according to Bulk API Limitations batch will fail if it has more than 10.000 records. So if user sets maximum records to something more than 10.000 we should tell the user that it does not make sense by failing the pipeline.

Salesforce Bulk API for INSERT and how we use it.

To insert records to Salesforce via Bulk API the following steps has to be taken.

1. Connect to Bulk API via OAuth2
2. Create a bulk job id
3. Split data into batches
4. Add batches to the job
5. Close job
6. Await every batch completion

7. Check results for every batch

In this section we will go though implementation details of every step.

STEP 1.  Connect to Bulk API via OAuth2

Using the provided clilentId, clientSecret, username and password,access token and instance URI will be fetched using username and password flow of OAuth2.

e.g:

grant_type=password&client_id=<your_client_id>&client_secret=<your_client_secret>&username=<your_username>&password=<your_password>

The following parameters are required:

Response would be :

{
"id":"https://login.salesforce.com/id/00D50000000IZ3ZEAW/00550000001fg5OAAQ",
"issued_at":"1296509381665",
"instance_url":"https://na1.salesforce.com",
"signature":"+Nbl5EOl/DlsvUZ4NbGDno6vn935XsWGVbwoKyXHayo=",
"access_token":"00D50000000IZ3Z!AQgAQH0Yd9M51BU_rayzAdmZ6NmT3pXZBgzkc3JTwDOGBl8BP2AREOiZzL_A2zg7etH81kTuuQPljJVsX4CPt3naL7qustlb"
}

This access token and instance URI will be used to execute the queries via bulk and soap api

STEP 2. Create a bulk job id

Ask Bulk API to create us a job and return it's id. So we can submit batches to it later.

The job type is set to insert, not an upsert. Since for upsert we need to know IDs for every record we update.

STEP 3. Split data into batches

Here we create multiple CSV files. Every of them contains multitude of records.

Splitting data into batches is done considering these 3 factors:

1. User configurations "Maximum bytes per batch" and "Maximum records per batch" must be obeyed.
2. Bulk API limitations must be obeyed. Here's the list of them:
   1. Batches for data loads can consist of a single CSV file that is no larger than 10 MB.
   2. A batch can contain a maximum of 10,000 records.
   3. A batch can contain a maximum of 10,000,000 characters for all the data in a batch.
   4. A field can contain a maximum of 32,000 characters.
   5. A record can contain a maximum of 5,000 fields.
   6. A record can contain a maximum of 400,000 characters for all its fields.
   7. A batch must contain some content or an error occurs.
   How we handle these:

   A,B,C - controlled by splitting the data correctly.
   E - checked during schema validation.
   D,F,G - if the data, which comes from sink, exceeds these, simply let the batch fail. Nothing we can do about these.

3. How many records comes to a specific mapper. For more details see <TODO>.

STEP 4. Add batches to CSV

Pass a FileInputStream of csv file to Bulk API and ask it to create a batch using this file.

STEP 5. Close Job

Ask Bulk API to close the job. This means that no more batches will be expected by Salesforce.

STEP 6. Await every batch completion

Poll for ever batch status until status is either Completed or Failed. Salesforce server enforces a timeout on batch processing, we don't need to implement any timeout logic here.

If any of batches fail, we fail the pipeline with an exception.

STEP 7. Check results for every batch

Query Bulk API to provide results of insertion of every single record. Result is a CSV. Every row of it represents a single record insertion result. Generally this will look like this:

Id,Success,Created,Error
fa4t2fggee,true,true,
rqewetrter,true,true,
,false,false,Field 'Name' is required and cannot be empty for sObject 'Contact'
gre3jvd245,true,true,

Records which has either success=false OR created=false are considered erroneous. Erroneous are processed according to user configuration "Error handling". For more information look at section "User configurations".

MapReduce Parallelization

BatchSink class uses NullWritable as key and CSVRecord as value.

BatchSink#initialize

Check which fields for given sObject are creatable by querying Salesforce SOAP API (for more information see "Policy regarding unknown or non-creatable fields")

BatchSink#transform

Create a CSVRecord instance using the StructuredRecord. In process convert logical types to expected format.

OutputCommitter#setupJob

Create Salesforce job, which the batches will be added to.

OutputCommitter#commitJob

Close Salesforce job.

RecordWritter#constructor

Create a tmp folder and establish connection to Salesforce Bulk API

RecordWritter#write

Append a record to a csv file. If according to the our batching policy we need the record to go into the new batch, we close the file and submit it to the Salesforce job as a separate batch. After that the new file is created and the record is appended to it.

For information on how we calculate batches please see Splitting data into batches.

RecordWritter#close

1. Close currently opened csv file and submit it as a batch to Salesforce job.
2. Wait for completion of EVERY batch which was submitted by current mapper.
3. Check results for every record in EVERY batch (submitted by current mapper), and act on them according to the error handling strategy configured by user.

Other points

Policy regarding unknown or non-creatable fields

Some points on behavior if schema contains fields which are not present or are not editable (creatable) in target sObject.

Let's consider a case where user wants to copy Contacts from one Salesforce instance to another. This is done by simply connecting Salesforce Source and Sink in ETL.

But this will inevitably fail, the reason for this is that any sObject contains a lot of fields which cannot be inserted like Id, isDeleted, LastModified and a lot of other fields which are often auto-generated.

These fields are different for every sObject. Good news is that we can query Salesforce SOAP API and check if any field in sObject is creatable.

Based on above said, I propose that we skip any non-creatable fields or fields that do not exist. We simply produce a log message with the list of fields that were ignored.

Converting fields

We will have to convert logical types like date, datetime, time from long to string format accepted by Salesforce. Other types won't require converting.

Compound fields

For source plugins we don't use nested records for compound fields. Instead we ask user to provide sub-fields for every compound object separately.

Compound fields %s cannot be fetched when a SOQL query is given.
Please specify the individual attributes instead of compound field name in SOQL query.
For example, instead of ‘Select BillingAddress ...’, use ‘Select BillingCountry, BillingCity, BillingStreet ...

I think handling of these fields should be consistent across all Salesforce plugins. Meaning sink needs to fail with similar error message, if somebody tries to insert a compound field directly. The message would state that user should fully qualify every sub-field he wants to insert.