Kusto ML plugins - CPU/Disk costs show 0 - azure-data-explorer

Queries that use KustoML plugins (e.g. autocluster/diffpatterns) take a while to run, but the CPU/Disk stats don't reflect this - they show 0.
Thus these queries don't show up in any analysis done to see what's using the most resources on a cluster, and also doesn't give the right feedback to users on the cost of these queries.

Related

Ingesting Google Analytics data into S3 or Redshift

I am looking for options to ingest Google Analytics data(historical data as well) into Redshift. Any suggestions regarding tools, API's are welcomed. I searched online and found out Stitch as one of the ETL tools, help me know better about this option and other options if you have.
Google Analytics has an API (Core Reporting API). This is good for getting the occasional KPIs, but due to API limits it's not great for exporting great amounts of historical data.
For big data dumps it's better to use the Link to BigQuery ("Link" because I want to avoid the word "integration" which implies a larger level of control than you actually have).
Setting up the link to BigQuery is fairly easy - you create a project in the Google Cloud Console, enable billing (BigQuery comes with a fee, it's not part of the GA360 contract), add your email address as BigQuery Owner in the "IAM&Admin" section, go to your GA account and enter the BigQuery Project ID in the GA Admin section, "Property Settings/Product Linking/All Products/BigQuery Link". The process is described here: https://support.google.com/analytics/answer/3416092
You can select between standard updates and streaming updated - the latter comes with an extra fee, but gives you near realtime data. The former updates data in BigQuery three times a day every eight hours.
The exported data is not raw data, this is already sessionized (i.e. while you will get one row per hit things like the traffic attribution for that hit will be session based).
You will pay three different kinds of fees - one for the export to BigQuery, one for storage, and one for the actual querying. Pricing is documented here: https://cloud.google.com/bigquery/pricing.
Pricing depends on region, among other things. The region where the data is stored might also important be important when it comes to legal matters - e.g. if you have to comply with the GDPR your data should be stored in the EU. Make sure you get the region right, because moving data between regions is cumbersome (you need to export the tables to Google Cloud storage and re-import them in the proper region) and kind of expensive.
You cannot just delete data and do a new export - on your first export BigQuery will backfill the data for the last 13 months, however it will do this only once per view. So if you need historical data better get this right, because if you delete data in BQ you won't get it back.
I don't actually know much about Redshift, but as per your comment you want to display data in Tableau, and Tableau directly connects to BigQuery.
We use custom SQL queries to get the data into Tableau (Google Analytics data is stored in daily tables, and custom SQL seems the easiest way to query data over many tables). BigQuery has a user-based cache that lasts 24 hours as long as the query does not change, so you won't pay for the query every time the report is opened. It still is a good idea to keep an eye on the cost - cost is not based on the result size, but on the amount of data that has to be searched to produce the wanted result, so if you query over a long timeframe and maybe do a few joins a single query can run into the dozens of euros (multiplied by the number of users who use the query).
scitylana.com has a service that can deliver Google Analytics Free data to S3.
You can get 3 years or more.
The extraction is done through the API. The schema is hit level and has 100+ dimensions/metrics.
Depending on the amount of data in your view, I think this could be done with GA360 too.
Another option is to use Stitch's own specfication singer.io and related open source packages:
https://github.com/singer-io/tap-google-analytics
https://github.com/transferwise/pipelinewise-target-redshift
The way you'd use them is piping data from into the other:
tap-google-analytics -c ga.json | target-redshift -c redshift.json
I like Skyvia tool: https://skyvia.com/data-integration/integrate-google-analytics-redshift. It doesn't require coding. With Skyvia, I can create a copy of Google Analytics report data in Amazon Redshift and keep it up-to-date with little to no configuration efforts. I don't even need to prepare the schema — Skyvia can automatically create a table for report data. You can load 10000 records per month for free — this is enough for me.

BigQuery bill breakdown? [duplicate]

Google Cloud billing is not updating with the free trial (on monthly payments) and I can not change it to a faster update cycle. As per https://cloud.google.com/free-trial/docs/billing-during-free-trial the bill should come every month.
It is therefore not easy to see how much of the 300$ is left.
Is there any way to at least see how many TBs my queries used? This should be by far the biggest item on the bill.
I am concerned that I might get 'stuck' between some important queries that I otherwise could have managed better to have at least partial results available after the trial ends.
BigQuery analysis & storage costs should be listed under your GCP billing transactions:
https://console.cloud.google.com/billing/<INSERT_YOUR_BILLING_ID_HERE>/history?e=13803970,13803205
Another way to see how much you have queried is by enabling audit logging as described here.

Run a DB-intensive query/calculation asynchronously

This question relates to WordPress's wp-cron function but is general enough to apply to any DB-intensive calculation.
I'm creating a site theme that needs to calculate a time-decaying rating for all content in the system at regular intervals. This rating determines the order of posts on the homepage, which is paged to allow visitors to potentially view all content. This rating value needs to be calculated frequently to make sure the site has fresh content listed in the proper order.
The rating calculation is not heavy but the rating needs to be calculated for, potentially, 1,000s of items and doing that hourly via wp-cron will start to cause problems for sites with lots of content. Ignoring the impact on page load (wp-cron processes requests on page loads once a certain interval has been reached), at some point the script will reach a time limit. Setting up the site to use "plain ol' cron" will solve the page loading issue but not the timeout one.
Assuming that I have no control over the sites that this will run on, what's the best way to handle this rating calculation on a regular basis? A few things that came to mind:
Only calculate the rating for the most recent 1,000 posts, assuming that the rest won't be seen much. I don't like the idea of ignoring all old content, though.
Calculate the first, say, 100 or so, then only calculate the rating for older groups if those pages are loaded. This might be hard to get right, though, and lead to incorrect listing and ratings (which isn't a huge problem for older content but something I'd like to avoid)
Batch process 100 or so at regular intervals, keeping track of the last one processed. This would cycle through the whole body of content eventually.
Any other ideas? Thanks in advance!
Depending on the host, you're in for a potentially sticky situation. Let me outline a couple of ideal cases and you can pick/choose where you need to.
Option 1
Mirror the database first and use a secondary app (WordPress or otherwise) to do the calculations asynchronously against that DB mirror. When they're done, they can update a static file in the project root, write data to a shared Memcached instance, trigger a POST to WordPress' admin_post endpoint to write some internal state, whatever.
The idea here is that you're removing your active site from the equation. The last thing you want to do is have a costly cron job lock the live site's database or cause queries to slow down as it does its indexing.
Option 2
Offload the calculation entirely to a separate application. Tracking ratings in real time with WordPress is a poor idea as it bypasses page caching and triggers an uncachable request every time a new rating comes in. Pushing this off to a second server means your WordPress site is super fast, and it also means you can have the second server do the calculations for you in the first place.
If you're already using something like Elastic Search on the site, you can add ratings as an added indexing facet. Then just update posts as ratings change, and use the ES API to query most popular posts later.
Alternatively, you can use a hosted service like Keen IO to record and aggregate ratings.
Option 3
Still use cron, but don't schedule it as a cron job in WordPress. Instead, write a WP CLI routine that does the reindexing for you. Then, schedule real cron jobs to process the job.
This has the advantage of using PHP's command line version, which can be configured to skip the timeouts and memory limits imposed on the FPM/CGI/whatever version used to serve the site. It also means you don't have to wait for site traffic to trigger the job - and a long-running job won't block other cron events within WordPress from firing.
If using this process, I would set the job to run hourly and, each hour, run a batch of 1/24th of the total posts in the database. You can keep track of offsets or even processed post IDs in the database, the point is just that you're silently re-indexing posts throughout the day.

Could "filling up" Google Analytics with millions of events slow down query performance / increase sampling?

Considering doing some relatively large scale event tracking on my website.
I estimate this would create up to 6 million new events per month in Google Analytics.
My questions are, would all of this extra data that I'm now hanging onto:
a) Slow down GA UI performance
and
b) Increase the amount of data sampling
Notes:
I have noticed that GA seems to be taking longer to retrieve results for longer timelines for my website lately, but I don't know if it has to do with the increased amount of event tracking I've been doing lately or not – it may be that GA is fighting for resources as it matures and as more and more people collect more and more data...
Finally, one might guess that adding events may only slow down reporting on events, but this isn't necessarily so is it?
Drewdavid,
The amount of data being loaded will influence the speed of GA performance, but nothing really dramatic I would say. I am running a website/app with 15+ million events per month and even though all the reporting is automated via API, every now and then we need to find something specific and use the regular GA UI.
More than speed I would be worried about sampling. That's the reason we automated the reporting in the first place as there are some ways how you can eliminate it (with some limitations. See this post for instance that describes using Analytics Canvas, one my of favorite tools (am not affiliated in any way :-).
Also, let me ask what would be the purpose of your events? Think twice if you would actually use them later on...
Slow down GA UI performance
Standard Reports are precompiled and will display as usual. Reports that are generated ad hoc (because you apply filters, segments etc.) will take a little longer, but not so much that it hurts.
Increase the amount of data sampling
If by "sampling" you mean throwing away raw data, Google does not do that (I actually have that in writing from a Google representative). However the reports might not be able to resolve all data points (e.g. you get Top 10 Keywords and everything else is lumped under "other").
However those events will count towards you data limit which is ten million interaction hits (pageviews, events, transactions, any single product in a transaction, user timings and possibly others). Google will not drop data or close your account without warning (again, I have that in writing from a Google Sales Manager) but they reserve to right to either force you to collect less interaction hits or to close your account some time after they issued a warning (actually they will ask you to upgrade to Premium first, but chances are you don't want to spend that much money).
Google is pretty lenient when it comes to violations of the data limit but other peoples leniency is not a good basis for a reliable service, so you want to make sure that you stay withing the limits.

total registered vs. concurrent users

Is there a proper way, equation or technique in general to say, "My web application needs to support N number of total users which via this equation/technique/rockHardExperience tells me that I need to support X number of concurrent page requests"?
From my research and/or gut feeling it seems like it would be something like:
totalLoadCapabilityRequired = (totalUsersN x .10 ) * .5
where .10 is for roughly 10% on at any given time
and the whole thing multiplied by 50% to suggest a 50% chance of those total users online executing a request at roughly the same time
any insights would help me in making sure I implement support in my application that is on par for the demand. I expect a lot of users but don't want to over anticipate too early. I know for starters that the org I am programming for will have 45,000 users that they want to use my system, with an anticipation on success for many more.
Here's a couple of things to think about:
What's the time span in which you expect the bulk of your visits? If it's an office application within the same physical company your capacity planning should be based on an 8 hour period. If most visits will come from the same continent you can plan for a 12 hour period instead, etc. Base your visitor spread on that.
Which pages do you anticipate will be the most popular and how heavy are those pages (i.e. how many pages can you load in one second)? Get an understanding of parts that would benefit from caching to squeeze out more performance.
Don't plan based on peak load; design your app to scale and start small.
Design your app in a way that you can take run snapshots at every 500th request; you can use tools like xhprof to create files that you can run through cachegrind tools to analyze the performance as it runs.
In short, there's no catch-all formula :) for a ballpark figure your formula will probably be good enough, but take the above points in consideration.

Resources