Help

Re: Filtering with Loops in Javascript - How to Avoid Insanity

5503 2
cancel
Showing results for 
Search instead for 
Did you mean: 

As developers, we get roped into codifying a vast and seemingly endless escalation of conditional processing to find and filter records and related tables. Where does it end? Apparently, it doesn’t because as soon as you create the perfect script that loops across many filtering constraints, the client (or management) is going ask for a “slight” modification or you will invariably discover an unanticipated case where you’re standing in a corner holding a paintbrush without egress.

This and vastly many of our script projects are fundamentally based on two essential tasks -

  1. Filter out the data that we need
  2. Perform computations on the filtered data

Typically, we use loops to achieve #1. #2 is easy if we get #1 right. But #1 is never simple and rarely performant given enough conditional complexity and records.

What’s better than this approach? Fewer loops for sure because they are slow, and they make our lives shi**y.

The Remedy…

The term “filter”, and especially in a javascript sense, is subject to abuse. Why not exchange this term with “search”. Indeed, a JSON hash index or an actual inverted index is exactly the right chainsaw to mitigate the looping and increase direct access to lists of record IDs that match what the loop is intended to provide.

I use hash indexes all the time, but increasingly, I use Lunr to establish inverted indexes that are very agile and extremely fast. Imagine finding 200 record IDs from a collection of 30,000 rows where three fields match specific values and the fourth is “fuzzy”. Looping requires lots of code and plenty of time. An inverted index can do this in 20 milliseconds.

And when you factor in the advantages of term presence, wildcard searches, and field-specific boosting, the loop that builds the index is time extremely well spent. Plus, you can create and cache-forward index documents independent of your business logic.

One last comment concerning perfomance - there are many computations and metrics that we perform repeatedly for business analytics. Why not compute these metrics and store them in the search index so that when the CEO wants to see how her business is performing, you’ve reduced the overhead to a simple search query rendered at the edge in 500ms?

While an inverted full-text index offers many advantages, there’s a simple approach you can use to improve script performance while creating an elegant approach that drastically simplifies your code when nested loops seem the only viable route.

Hash Index

This is simply a JSON document intended to make it fast and simple to perform lookups; it’s literally a copy of a collection of items where the object key represents a way to directly access the data related to that key.

Let’s say you had a thousand records in a table and you needed to loop through it to find a given record whose name matched a specific customer name – and you need to perform this looping lookup repeatedly nested inside another enumeration. You have three options -

  1. Perform a FilterByFormula approach for each lookup; ugly, slow, convoluted.
  2. Loop through every record seeking a match; very slow and CPU-intensive, added code complexity.
  3. Perform a “seek” into a pre-built hash index; almost instant, elegant.

Approach 3 requires a single pass across all the records containing the customer names (that’s the downside). That pass would create a collection of objects in a single JSON document (ex., oCustomerNameHash that is global to your app) and looks something like this:

{ "ABC Transportation" : {
   "customer_id" : "1001",
   "customer_contact" : "Jimmy Johns",
   other attributes needed for the index...
   }
}
... objects representing rest of the customers ...

When you want to know the contact name for “ABC Transportation”, the code is simple. effortless, and requires only 6 milliseconds.

let customerContact = oCustomerNameHash["ABC Transportation"].customer_contact;

Even if the hash index has 50,000 items in it, the response time will be sub-20 milliseconds. This single line of code eliminates a nested loop and the complexity of doing so. It is possible because you performed the loop in advance of needing to perform the lookups. Essentially, you are caching forward an intelligent and performant class of data objects that can be reused throughout your script.

This makes it possible to vastly compress the code and complexity while also sparing the number of loop introspections required to a single pass across that data instead of vastly many passes which are required in nested filtering loops.

With a hash index like this, you may often find multiple cases where the hash index can be utilized, so it has the advantage of being ready to perform lookups for the duration of your script process. Furthermore, you may discover you need to perform lookups using additional keys which suggests multiple hash indexes for the same data.

Imagine you wanted to also lookup customer name by customer ID or customer contact by customer ID. This would require a new hash index like oCustomerIDHash. Instead of building a second hash, this added seek capability can be built into the same constructor loop so there’s no need to loop for every index dimension you might need.

Example… you can mix both indexed dimentions into one hash index document - this time we’ll call it oCustomerHash. By simply creating two keys for each customer - one for the customer name, and one for the customer ID – we have begun to create a complex indexing model. The lookup process is constant and the agility is magnified.

The objects in the index would include two entries for each customer.

{ "ABC Transportation" : {
   "customer_id" : "1001",
   "customer_contact" : "Jimmy Johns",
   other attributes needed for the index...
   }
}

{ "1001" : {
   "customer_name" : "ABC Transportation",
   "customer_contact" : "Jimmy Johns",
   other attributes needed for the index...
   }
}

Now you can query by customer name to get customer ID or by customer ID to get customer name instantly - no nested looping required.

let customerName = oCustomerHash["1001"].customer_contact;
let customerID   = oCustomerHash["ABC Transportation"].customer_id;

Building the Index

This is one approach that can be used to construct the index from an existing collection of records.

let oCustomerHash = {};
for (let record of oRecords)
{
    // add the customer name object
    oCustomerHash[oRecords[record].getCellValue("Customer Name")] = {
        "customer_id"      : oRecords[record].getCellValue("Customer ID"),
        "customer_contact" : oRecords[record].getCellValue("Customer Contact"),
     }
    // add the customer id object
     oCustomerHash[oRecords[record].getCellValue("Customer ID")] = {
        "customer_name"    : oRecords[record].getCellValue("Customer Name"),
        "customer_contact" : oRecords[record].getCellValue("Customer Contact"),
    }
}
31 Replies 31

It depends on the nature of the process. If an update process depends on lots of lookups, it’s likely a hash index will be beneficial.

I would add these finer points to use cases where hash indexes are helpful.

  • When you need to perform lookups repeatedly. The big dividend for hash indices is paid out when the index can be used dozens, hundreds, or even tens of thousands of times.

  • Finding records through a blistering fast index key is a common use case, and it’s deeply beneficial because you only need to load the data once despite needing hundreds o thousands of possible seeks into that data. However, hash indexes are useful for all sorts of tabular data, not just Airtable records. Imagine loading the entire US counties database over an HTTP request and then performing lookups to gather data properties from the external counties database. A hash index allows you to instantly cross reference by a simple post-code lookup that takes a few milliseconds.

@kuovonne Yes, this is what happened. I realized after that I was saving each one. I changed it to batching shortly after I asked the above question and it’s much quicker. If I do a Scripting Extension, am I still capped at updating 50 records at a time? And I’m assuming there’s no way to trigger a scripting extension from an automation? Or have these run automatically?

@Bill.French Thanks! I’ll certainly be keeping this one in my back pocket. Seems super useful.

Yes, you are still capped at sending 50 records in a single request, and 15 requests per second.

Both scripting extension and scripting actions can run scripts. You cannot call scripting extension from an automation, but you can often adapt a script from one environment to run in the other.

Since you switched to batching updates, you should have no problem keeping the script as an automation script.

Yes, and when doing this, it leaves few other dials to increase performance which means, higher throughput must come from all that happens around the 50-item batching.

One approach to faster performance is to not need fast throughput in the first place. You must assess things like -

  • Is it possible to do this process slowly without impacting users or the usability of your solution?
  • Reflect on the pathway that got you into this challenge in the first place? Is the data model painting you into a corner where massive batch updates are necessary?
  • Have you exhausted the possibility of a fully event-driven [real-time] architecture? This is to suggest that by processing data items as they occur, you spread the updates over minutes, seconds, and micro-seconds to eliminate the need to ever batch process anything.
Nick_Richards
5 - Automation Enthusiast
5 - Automation Enthusiast

@Bill.French Interesting questions that I need to ask myself.

So I have a bunch of companies that I do specific tasks for. Every quarter, I do the same task and process for each customer. I use Airtable to track the status during these quarterly tasks. Each quarter, I want to reset the status for each company to the beginning of the process.

So, is there a better way to do this reset? I don’t typically mass update records. But I love hearing how others tackle problems.

Excellent points. I was thinking of where hashes are most useful in the context of an Airtable script.

Given that @Nick_Richards says that his script is 8 lines long, updates only 150 records, and runs only once a quarter, I think that it is pretty safe to say that he does not need to implement the excellent techniques that you are suggesting.

If you want to continue with scripting, this is definitely a technique to keep handy. Bill imparted it to me early in my Airtable scripting journey and it has served me well. However, it has been many years since Bill was new to coding, and some of his advanced techniques are for more extreme situations than new coders typically encounter.

LOL! Then this should take about 9 seconds max unless I’m missing something.

You’re over thinking it. He was originally updating records one at a time in an automation script.

And nine seconds is a bit long to me. I would expect more like 3-5 seconds (most of which is spent waiting for the three batches of 50 records to be updated).

My comment was intended in the context of an Airtable script. We often assume that to use external data we must import it into Airtable and then lookup into it with a script or other relational automation process. Better, in some cases, to avoid bringing in data and relying (instead) on late binding processes that perform very fast queries against a hash index.