Joseph M.

In an ideal world, we should not have the concept of warehouses in Snowflake. I don’t think I have ever met a data engineer who wants to delve deep into warehouse internals when they are already juggling tasks from different business stakeholders. Ideally, warehouses in Snowflake or DBUs in Databricks are just units of compute. One should be able to define what needs to be done and the underlying data warehouse should be capable enough to figure out the size of machine and the parallelism it needs to complete my task within a particular timeframe and minimum cost. No one cares about the configuration of the warehouse if the task succeeds within the business SLA and it’s priced fairly. Though to be honest, accurately predicting the memory and runtime requirements of an SQL before actually executing is a very tough technical problem to solve. Snowflake has always been famous for ‘it just works’ and my gut says that it’ll be the first to solve this problem (if it isn't solved by academia), but for now, it remains unsolved and the users need to take care of this! 

49

13 Comments

Someone rightly said : Snowflake is for rich, low to no-tech departments like Finance. Databricks is for data engineers. In the end, both are successful, multi-billion dollar companies. Choice in the end lies in defining a clear customer segment. #snowflake #databricks #dataengineering #customersegment

41

3 Comments

I am back after the Diwali vacation with my quirky posts about Data Engineering. Ever wondered what’s happening under the hood of a PySpark cluster? 🤔 Well, buckle up because I’m about to demystify it all in this post! 🎬 𝗧𝗵𝗲 𝗣𝘆𝗦𝗽𝗮𝗿𝗸 𝗠𝗼𝘃𝗶𝗲 𝗦𝗲𝘁 🎬 𝗗𝗿𝗶𝘃𝗲𝗿: The director 🧑‍🎤. Calls the shots, decides on the overall plan, and ensures everything runs smoothly on set. 𝗖𝗹𝘂𝘀𝘁𝗲𝗿 𝗠𝗮𝗻𝗮𝗴𝗲𝗿: The production manager 📋. Keeps track of all the moving pieces, ensuring resources and personnel are allocated where needed. 𝗪𝗼𝗿𝗸𝗲𝗿 𝗡𝗼𝗱𝗲𝘀: The film crew 🎥. Each team handles a different part of the production process—like lighting, sound, and camera work. 𝗘𝘅𝗲𝗰𝘂𝘁𝗼𝗿𝘀: The crew members 🎞️. These folks carry out specific tasks, like setting up scenes, managing props, or controlling the camera. 𝗧𝗮𝘀𝗸𝘀: Individual shots 🎬. Each shot must be perfectly timed and executed to ensure the final film looks fantastic. 𝗣𝗮𝗿𝘁𝗶𝘁𝗶𝗼𝗻𝘀: Scenes from the script 📜. Each scene is shot independently but must come together to form a cohesive story. 🎞️ 𝗔𝗰𝘁𝗶𝗼𝗻! Lights, camera, action! The Driver (director) calls the shots, ensuring the film crew (worker nodes) and crew members (executors) all align to get the perfect shot (tasks) for each scene (partitions). And just like that, your blockbuster data processing movie is a wrap! #PySpark #DataIsArt #BigDataProduction #ClusterManagement #DataEngineering #TechAnalogies Ready for the next take? 🎬🎥📽️

21

**𝐋𝐞𝐯𝐞𝐥 𝐔𝐩 𝐘𝐨𝐮𝐫 𝐃𝐚𝐭𝐚 𝐒𝐤𝐢𝐥𝐥𝐬 𝐰𝐢𝐭𝐡 𝐒𝐧𝐨𝐰𝐟𝐥𝐚𝐤𝐞 𝐓𝐫𝐚𝐢𝐧𝐢𝐧𝐠!** A huge shoutout to my friends, Vishal Kaushal , a Snowflake Squad member, and Sachin Mittal, a 4x Snowflake Data Superhero, who are starting Snowflake Training: Zero to Hero course starting May 18th! . **Why Snowflake? ** Snowflake is more than just a data warehouse. It's a powerful, cloud-native data platform that empowers data engineers to build next-generation data solutions. Here's why learning Snowflake can be a game-changer for your career. By mastering Snowflake, you'll equip yourself with the skills to design and implement robust, scalable data solutions that are in high demand across industries. This positions you as a valuable asset in the ever-evolving world of data engineering. This comprehensive training, will equip you with the knowledge and skills to become a Snowflake master. Here's what you'll gain: Solid Foundation: Grasp the core concepts and functionalities of Snowflake. Hands-on Learning: Get practical experience through real-world exercises and projects. Unlimited doubt support Real time Projects : 2 End to End Real time projects , 1 using Matillion as well. Ready to take your data engineering skills to the next level? Don't miss out on this amazing opportunity! Sign up for the Snowflake Training: Zero to Hero course starting May 18th. 👉🏻 Enroll Here - https://lnkd.in/gcfxzNx4 📅 Live classes starts from 18th May 2024 Discount Coupon : LIVE20 PS : This is not a paid post. #Snowflake #DataEngineering #CloudDataWarehouse #Training #Upskilling #DataSuperhero

121

4 Comments

In this video, we look at the internals of Facebook's graph datastore. This system, called "The Associations and Objects (T.A.O)", is used across Facebook, Instagram and WhatsApp to represent objects and relations in memory. The graph is built over Memcached. Facebook engineers store petabytes of data in this graph and have scaled it to 1 billion queries per second. I have covered this system in detail at InterviewReady during live classes. Glad to see the enthusiasm for research papers :D Cheers! #SystemDesign #Facebook #Graph https://lnkd.in/dGdM-nb7

49

Databricks surely gonna be ruling the Data and AI space. A right step towards right direction at right time 💥 Databricks has acquired Tabular, a data management startup founded by Apache Iceberg creators, for over $1 billion. The acquisition aims to build a unified data management platform based on open lakehouse formats, reducing silos and friction for customers. This move highlights Databricks' commitment to open formats and open source data in the cloud. Read the detailed news over here: https://lnkd.in/gqVcxjcv #dataengineering #dataplatform #datastrategy #storage #nosql #cloud

6

𝗦𝗲𝗺𝗶-𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲𝗱 𝗱𝗮𝘁𝗮 𝗱𝗼𝗲𝘀𝗻'𝘁 𝗵𝗮𝘃𝗲 𝘁𝗼 𝗯𝗲 𝘁𝗼𝗼 𝗵𝗮𝗿𝗱   Working with varied data formats that have complex or unclear structures is often challenging. Today's data ecosystem deals with more than just traditional rows and columns, with modern databases needing to process complex formats like JSON, XML, and Parquet.   Semi-structured data provides the flexibility to work with dynamic data without requiring a rigid schema. JSON has become popular because it's easy to read, lightweight, and works well for sharing data between web apps, APIs, and databases.   Snowflake has features that make working with JSON easier and faster, with less hassle and more insights.    To help with this, here's a 7,000-word guide on how to query JSON data in Snowflake. ⤵️   Comment what you'd like to learn about MongoDB and JSON data next, and we'll publish it. 😁 #dataingestion #blog #guide #snowflake #query #json #data #datazip #database

14

I recently began dashboarding on Retool, with data sources being RDS Postgres and Bigquery. I've always used Python+ SQL for data manipulation. After some time, I realized that performing SQL queries for each data piece would result in repetitive calls to resources. Retool introduced me to the concept of JS Transformers. Transformers allow us to construct larger, reusable sections of JavaScript code that can be accessed across the program, reducing the need to perform several SQL queries. Instead, write a single SQL query and then utilize transformers to extract the relevant data. However, using these transformers requires a learning curve, and I began to like Reduce Methods. Resource Links: Reduce Methods in JS: - https://lnkd.in/dbF_ybam - https://lnkd.in/dzF_n9-F JS Transformer: https://lnkd.in/d9cAzpsU image credit - Matt Lewandowski

35

3 Comments

If you're new to Databricks or are looking to further develop your skillset, the Databricks academy is a great place to start. Here you can learn about Databricks sales and technical training to get enabled on the Data Intelligence Platform. Before you can help your customers unlock the potential of their data, you need to master the Databricks Lakehouse Platform. The Databricks Academy is your go-to place for training and certification for all roles in your organisation. If you would like to discuss the resources available and where to begin, feel free to reach out and I can point you in the right direction. #learning #certified #databricks

23

Spark Streaming!! Interviewer: "Today, let's explore Spark Streaming, an integral component of Apache Spark for processing real-time streaming data. Are you familiar with Spark Streaming?" Candidate: "Yes, I've worked with Spark Streaming. It provides a scalable and fault-tolerant framework for processing real-time data streams." Interviewer: "Excellent! Spark Streaming enables developers to build robust and scalable stream processing applications. Could you elaborate on the key features and benefits of Spark Streaming?" Candidate: "Certainly! Spark Streaming offers several advantages. Firstly, it provides high throughput and low latency processing of real-time data streams. Secondly, it integrates seamlessly with the core Spark API, enabling users to leverage Spark's rich ecosystem of libraries and tools. Additionally, Spark Streaming offers fault tolerance through lineage and checkpointing, ensuring that data processing is resilient to failures. Lastly, it supports various data sources and sinks, allowing users to ingest data from sources like Kafka, Flume, or HDFS, and output processed data to databases, file systems, or dashboards." Interviewer: "Well summarized! Now, let's delve into the architecture of Spark Streaming. How does Spark Streaming process and analyze real-time data streams?" Candidate: "Spark Streaming follows a micro-batch processing model, where real-time data streams are divided into small batches of data, which are then processed using Spark's batch processing engine. The architecture consists of a Receiver that ingests data from input sources, a StreamingContext that manages the execution of streaming jobs, and a Discretized Stream (DStream) abstraction that represents a continuous stream of data. Each batch of data is processed using RDDs (Resilient Distributed Datasets), enabling parallel and fault-tolerant stream processing." Interviewer: "Spot on! Spark Streaming's micro-batch processing model provides fault tolerance and scalability, making it suitable for handling large-scale streaming data. Now, let's discuss the use cases of Spark Streaming. Can you provide examples of scenarios where Spark Streaming excels?" Candidate: "Certainly! Spark Streaming is well-suited for various real-time analytics and monitoring applications. For example, in e-commerce, it can be used to analyze user activity and recommend products in real-time. In finance, it can process streaming market data to detect anomalies or fraud in transactions. Additionally, in telecommunications, it can analyze network logs to identify network issues or optimize network performance in real-time." Interviewer: "Impressive examples! Spark Streaming's versatility makes it applicable across different industries for real-time data processing and analysis. Thank you for sharing your insights on Spark Streaming!" Candidate: "You're welcome! #bigdata #spark #sparkstreaming #pyspark #python

7

🔍 Analytics Engineers, are you focusing on data modeling? 💡 Data leaders regularly tell us that a lack of or insufficient data modeling skills is one of the key blockers when hiring Analytics Engineers. Why is this? 🤔 📉 Has dbt lowered the bar? 🏢 Are certain companies not prioritizing a formal data modeling approach, resulting in a shortage of data modeling skills? 📚 Are more junior data folks not investing enough time in learning this skill set? Below are some links to The Stacked Data Podcast where data modeling has been part of the conversation 🎧: 🎙️ Andrea Salvati - https://lnkd.in/edqnUahx 🎙️ Rob Winters - https://lnkd.in/e2AYsVqR 🎙️ Thomas in't Veld - https://lnkd.in/eYHBDT3M If you are reading this and data modeling is a key component of your skill set, please get in touch as we have some brilliant Analytics Engineering opportunities! 🚀✨ Cognify Search

17

Snowflake Optimization - Summary of the week 1. Remove failing workloads Failing workloads are just using credits and not performing any real work. To identify them, use this: https://lnkd.in/gP8bgV-a 2. Identify unused storage Identify and remove unused storage. Here's a blog to identify: https://lnkd.in/gMTrHxmM 3. Set query timeouts Set appropriate query timeouts based on the size of the warehouse and the business use case. 4. Split or merge files before ingestion The appropriate file size for ingestion into Snowflake is 100-250MB. Split or merge files before ingesting to ensure the file size is correct. 5. Use deferred merge If the frequency of ingestion is high, use the concept of deferred merge. Here's a blog: https://lnkd.in/gk4jgz2g

20

3 Comments

One of the perks of becoming a "Know nothing influencer" Folks may ask you to review their books! Vinishka Kalra blessed me with an advance copy of "Databricks ML in Action", which I was reading up here on the beautiful shores of Lake Superior this week. It's so good. So who is the book for? I'd say it's for intermediate level ML practitioners who are new to Databricks. I'm a data engineer, so I spend my time *not* doing ML so there were new concepts here in this book for me. Those not as familiar with how to work with ML models, and the MLOps thing may want to do some pre-study before diving in to gain the full value of the text. This would be the "ML model" serving use case as outlined in the Data Engineering Lifecycle promulgated by the great Joe Reis 🤓 for other Data Engineers out there I think. Also, if you have been around Databricks before, as I have, there may be some more review here for you in the beginning. Now to gush. The book is a fantastic tour in *specific* *practical* examples interrelated with the concepts the authors are trying to introduce. Your mission with this tome, should you accept it, will be to produce four different projects integrating various features of the Databricks platform. You'll make: 1. A time series forecaster 2. A Feature engineered stream of transactions 3. A multi-label image classifier 4. (My personal favorite) A retrieval augmented generation equipped chatbot. Each of these projects you will develop code snippet by code snippet, across each of the chapters directly in line with the concept the authors teach. An excellent way of teaching I think. I wonder if Zach Wilson's LLM grader used RAG for the 4th iteration of his awesome bootcamp (which you should totally get as well as the book).

20

11 Comments

Real-Time Indexing at Scale: How Meesho Leveraged Apache Flink 🚀 At Meesho, they support millions of entrepreneurs across Bharat. With over 50 million products in their catalog, keeping product data fresh in real-time is essential. Delays in indexing lead to users seeing outdated prices or out-of-stock items, causing frustration, cart abandonment, and potential customer loss 😯 Here's how we overhauled our indexing architecture using Apache Flink for horizontal scalability 👇🏻 They started with a crontab-based approach, scheduling indexing jobs every 10 minutes. It worked well initially, with indexing taking 5–7 minutes for our smaller product base 🕑 The Scaling Challenge 📈 As Meesho grew, so did their problems 👇🏻 1️⃣ Exponential Growth: The number of products exploded into tens of millions, pushing indexing time to 4 hours. 2️⃣ Data Lag: Users saw outdated prices and stock information, especially during peak sales. 3️⃣ Processing Inefficiency: Recomputing all product data—even if unchanged—added unnecessary load. 4️⃣ Failure Delays: Job failures caused delays of up to 8 hours. Interim Fix: Kafka Consumers ✅ While plotting a long-term solution, they implemented Kafka consumers to 👇🏻 1️⃣ Parallelize Elasticsearch Writes: This reduced ES write time from 2 hours to 15 minutes, cutting overall indexing time to ~3 hours. 2️⃣ Adopt Batch Processing: Catalogs were sent directly to Kafka for immediate indexing, reducing memory consumption. Though this helped temporarily, as their catalog grew, indexing time soon crept back up to 4 hours. Requirements for the New Solution ✍🏻 They needed a system that was 👇🏻 1️⃣ Low-latency & fault-tolerant: Processing events in seconds, not hours. 2️⃣ Decoupled: Separating indexing from the enrichment pipeline. 3️⃣ Extensible: Supporting multi-consumer queues and future data flow operations. Why They Chose Apache Flink ❓ After evaluating several options, Apache Flink stood out for 👇🏻 1️⃣ Exactly-once processing guarantees: Ensuring data consistency. 2️⃣ Very low latency: Even with fault tolerance enabled. 3️⃣ High throughput: Capable of processing 182 million records/sec across 30 nodes. Flink vs. Storm: Performance Benchmark 😎 1️⃣ Flink (no fault tolerance): 182M records/sec, 0 ms latency. 2️⃣ Flink (with 5 sec checkpoints): 178M records/sec, 0 ms latency. 3️⃣ Storm (no fault tolerance): 9.8M records/sec, 11 ms latency. 4️⃣ Storm (with fault tolerance): 0.57M records/sec, 30–120 ms latency. 5️⃣ Flink had 300× higher throughput than Storm, even with fault tolerance. ✅ Read Full blog - https://lnkd.in/gQS_WyQt 🚨 Join my high quality, industry & modern tech stack based and practical project driven Data Engineering BootCAMP✌🏻👇 👉 Enroll Here - https://bit.ly/4eA2tuX 🚀 Dedicated Doubt Support & Placement Assistance 📲 Call/WhatsApp for any query (+91) 9893181542 Cheers - Grow Data Skills 🙂 #dataengineering

78

2 Comments

One of the most common question asked in Data Engineering Interview: Difference between ETL, ELT, and Reverse ETL. 1️⃣ ETL: The Traditional Powerhouse Extract, Transform, Load (ETL) has been the go-to process for decades. It involves: 1. Extracting data from various sources 2. Transforming it to fit operational needs 3. Loading it into the target system (usually a data warehouse) 2️⃣ ELT: Flipping the Script Extract, Load, Transform (ELT) emerged with the rise of cloud computing and big data. The key difference: - Data is loaded into the target system before transformation - Leverages the power of modern data warehouses for transformation - Offers more flexibility and scalability 3️⃣ Reverse ETL: Closing the Loop A newer player in the field, Reverse ETL: - Moves processed data from warehouses back into operational systems - Enables data activation, turning insights into action - Bridges the gap between analytics and operations For career guidance in Data Analytics, job search strategies, resume review, interview preparation, etc., you can book a session with me here: https://topmate.io/itshb

53

Want to embed large corpora blazingly fast? ✅ Care about SOTA quality? ✅ And don't want to pay too much? ✅ Snowflake's arctic-embed family, is now available on HuggingFace under an Apache2 license 🥁 It's the most practical family of text embeddings in the world, outperforming models that are 4-20x the size and upto 4x the embedding dimensionality. Are you a hobbyist looking for a tiny model to deploy on your laptop? snowflake-arctic-embed-xs (23M) and snowflake-arctic-embed-m (33M) are your new best friends ... 💻 Building a search engine from scratch and looking to embed 100s of billions of docs and care for cost? Meet snowflake-arctic-embed-m and snowflake-arctic-embed-m-long ... 🔎 Looking for the right enterprise sweet spot between large corpora, high quality and cost efficiency? Our 334M param model snowflake-arctic-embed-l outperforms blackbox API models from providers like OpenAI Cohere and Google while being significantly smaller and faster. 🔥 Use them in LlamaIndex, LangChain and your custom RAG pipelines and enjoy!

53

1 Comment

Some issues on #Robyn MMM that I have noticed. Would love to hear if there is a fix around these: 1. Different folks getting different results on same code. I have noticed the same code producing different results from different analysts. Robyn's default code doesn't have the set.seed option but can be easily added. Despite that, results vary. Another code change is in Step 3: Build Initial model step. Despite that, I have seen results being inconsistent with model outputs. 2. The optimization tends to be sticky Even when we know there is room for marginal spend in certain channels, the optimization continues to show pretty sticky results in its allocation. 3. Contributions highly proportionate to spend In general, we do see spend being correlated to contribution, but the percentage of contribution and spend are almost 1 to 1 in a lot of cases. This is an issue if you are trying to see if we are overspending in a certain channel that drives a lot of sales... Would love to see if you have ideas to get around it.. What else have you encountered or solved for? #MMM #Robyn #meta

28

20 Comments

Do you 𝑟𝑒𝑎𝑙𝑙𝑦 need 𝐝𝐛𝐭? dbt is on 🔥 and every consulting company now claims to know dbt (not all are created equal, ask me 😄), but should your organization jump on the bandwagon? 1️⃣ Do you lack data documentation? 📖 While dbt won't write the documentation for you, it enables teams to capture docs and expose them to end users via a Data Catalog of choice. 2️⃣ Would adding data quality or unit tests to your process help you detect issues before your end users? ✅ There are many ways you can perform Data Tests and there are lots of open source and paid options, but the first step is to actually do something. dbt also has packages like 𝐝𝐛𝐭-𝐞𝐱𝐩𝐞𝐜𝐭𝐚𝐭𝐢𝐨𝐧𝐬 so you don't have to reinvent the wheel and now you can do unit tests in addition to data tests using just dbt. 3️⃣ Is it hard for you to pin point the source of a bad calculation or understand the impact of a change to a source? 📊 dbt provides a lineage graph and you can enrich it with exposures which link data in your data warehouse to dashboards, ML models, etc. With dbt, it is relatively straight forward to find the impact of a change and find the cause of a production issue. 4️⃣ Do you have "black box" transformation processes where users have to "trust" things were done "right" and different teams use different tools? 🧩 It is common to see tools for 𝐈𝐓 and tools for "business" users. Traditional tools may have GUI interfaces and give some people the false sense of simplicity, but the biggest issue is that you have multiple people doing the same type of work using different tools. I am not suggesting there is no need for Excel or other business friendly tools, but I am saying that the bulk of your data transformation should be done in a consistent way. 5️⃣ Do you manually validate changes and manually deploy code to production or worse yet, you only have a single production database where you make changes directly? 💥 I shouldn't have to tell you how bad it is to change production directly. A bad query can take down the database for everyone or worse yet, deletes some data unintentionally. With dbt you can have isolated development environments, perform automated tests in isolation, and deploy code to production without manual intervention. Today, no one should be making changes without first committing code to git and promoting to production. When you implement good 𝐃𝐚𝐭𝐚𝐎𝐩𝐬 processes you can also enforce governance rules by using tools like 𝐝𝐛𝐭-𝐜𝐡𝐞𝐜𝐤𝐩𝐨𝐢𝐧𝐭 and SQLFluff. dbt is not right for everyone and there are other options out there, but there are many companies that can benefit from implementing dbt. I won't sugar coat it, you also need to change your mindset, process development is not trivial, and platform setup and management can be a pain which is why we created Datacoves, but implementing dbt correctly will help many companies both large and small. Are your data pipelines robust?

105

13 Comments

Week 6 of Elite Data Engineering Master's Program Summary Sumit Mittal and TrendyTech Enforcing the Schema – Schema should not be inferred and it should be enforced due to below issues – 1.   Spark may not infer data types correctly. 2.   Spark needs to scan more data in order to infer the schema which hampers the performance and adds more load to Apache Spark. Scanning more data takes more time. Through samplingRatio we can ask Apache Spark to scan some part/percentage of data. There are two ways to enforce the schema – 1.   Schema DDL 2.   StructType How to deal with date types – Apache spark understands yyyy-MM-dd date format. If we give date in another format and do not specify the format, spark will throw and error. There are two ways to handle date formats – 1.   Specify the date format while reading the data frame. 2.   Load the date as string and later convert the data type to relevant form. Read Modes – There are three modes to read the data frame – 1.   failfast – 2.   permissive – 3.   dropmalformed – Different way of data frame creation – 1.   Standard data frame reader API – spark.read 2.   Shortcut methods to read the dataframe – spark.read.csv, spark.read.json, spark.read.orc, spark.read.parquet 3.   The output of spark.sql is always a data frame 4.   Spark Table/Spark SQL can be converted to data frame using spark.read.table/spark.table 5.   Spark.range() will give single column dataset for testing purpose 6.   spark.createDataFrame to create data frame from local list Converting RDD to data frame – Below are the ways to convert RDD to data frame. 1.   df = spark.createDataFrame(rdd).toDF(column_names) 2.   df = rdd.toDF(orders_schema) 3.   df = spark.createDataFrame(rdd,schema) Nested Schema – Nested schema can also be handled using SchemaDDL and StructType. Also, local lists having nested schema also can be converted to data frame using createDataFrame. Data Frame Transformations – select vs selectExpr To add new column or modify existing column – withColumn transformation To rename column – withColumnRenamed transformation To drop column – drop transformation selectExpr will accept the column names and expressions as well. We can create new column and update existing column using expressions in withColumn, select and selectExpr. Removal of Duplicates from Data Frame – .distinct is good when we are considering the duplicated across all the columns. dropDuplicates is best option when we want to remove duplicates with selected columns. Spark Session in Details – Spark Session is an entry point to spark cluster. Spark Session is needed for higher level APIs and RDD/Spark Core APIs use spark context as and entry point to the spark cluster. Spark Session holds other contexts (spark context, hive context, SQL context). #bigdatadeveloper #trendytech #bigdatatechnologies

61

5 Comments