I have read so many articles before on Spark Memory Management but this is the first time I have understood the allocation and other details so clearly. Thanks a lot. Really helpful video.
Hey Afique, thanks for the wonderful explanation. Ok, so now I have got a few questions, plz clear the doubts: 1. In the unified memory, what if the the execution memory needs the full space that is occupied by storage memory, can the blocks from the storage memory be evicted to make room for the execution memory? So, can the execution memory occupy 100% of the space of unified memory (execution + storage) 2. If yes, so let's suppose an event where the execution memory occupies the full unified memory and it still needs more memory. 3. So, in this case, we have two choices -a disk spill or an off heap memory, we should opt for off heap memory over disk spill as u explained in your video . 4. The most important question now is that if we can use disk spill or off heap memory why do we get Out Of Memory Error in executors. I hope that you got my points and will soon get clear explanations from your end. Thanks again.
Hi @vinitrai5020, Good question! Yes, execution can request 100% of space from the unified memory manager pool, however, in cases where you want to immune the cached blocks from eviction, you can always set `spark.memory.storageFraction` to a value. If you set this value to, for e.g. 0.1, 10% of the total memory cannot be evicted by execution. However, it is important to note that this is on-demand. If `spark.memory.storageFraction` is set to 0.1 (10%) but nothing is cached, execution will just go ahead and use that storage memory and storage will wait for that 10% memory to free up before it can use it. Refer Spark documentation here: spark.apache.org/docs/latest/tuning.html#memory-management-overview On Spark throwing OOM errors, despite always having the option to spill to disk is because most in-memory structures used for joins, aggregations, sorting, shuffling cannot be “split”. Consider an example where you’re doing a join or an aggregation. In this operation, the same keys land in the same partition. Imagine one of the join/aggregation key being so large that it doesn’t fit in-memory. Now, spill doesn’t work here because that in-memory structure “supposedly” holding that large key cannot be “split” i.e. depending on the nature of data, half of the join cannot be done while spilling the rest and then later getting the spilled data back and doing the join for this half. This is primarily because that in-memory structure cannot be “split”. Enabling off-heap memory would help reduce the memory pressure and now: - Total execution memory = execution (on-heap) + execution (off-heap) - Total storage memory = storage (on-heap) + storage (off-heap) If the size of the large key (as discussed above) is good enough to fit in the total execution memory after enabling off-heap memory, an OOM will be avoided. Hope this clarifies :)
Top Class brother! Simple, Amazing and impactful. You deserve great appreciation to bring these internals. May God bless you with great health, peace, mind and prosperity! Keep growing.
I spent INR 42000 on a Big Data course but could not understand this concept clearly because the trainer himself lacked clarity. Here I understood completely.
I have two accounts in RU-vid and subscribed in both, Reason is you are putting some serious effort into the content. Beautiful Diagrams clear explanation accurate information is beauty of your content. Thanks, Afaque Bhai
Thanks Afaque Ahmad, very good series and loved all of them. Good work I have a few questions for you, may be we can discuss here if possible or if you are planning a new video, I will wait for it. 1. Here you discussed about Executor Memory Management. What about Driver Memory Management, how it works internally? 2. What are the similarities between Executor and Driver Memory Management? 3. What are the differences between Executor and Driver Memory Management? Many thanks in advance.
Hey @rambabuposa5082, thank you for the kind words, really appreciate it :) Regarding `Driver Memory Management`, appreciate the ask, but I do not have plans yet for a video. Reason is, I believe Driver & Executor memory management go hand-in-hand and relatively easy to manage Driver if your concepts are clear on Executor memory management because of several similarities (as you asked in one of your questions). Internally their memory components look similar in the sense that they both have JVM (on-heap) and off-heap memory and the division/logic of memory in the driver is just the same as the executor. Key differences are in terms of "roles and usage". You would have 1 driver which is solely responsible for creating tasks, scheduling those tasks, communicating back and forth with the executors on progress and aggregating the results (if needed), therefore its memory usage patterns differ from those of executors, which perform the actual data processing and storage. An important difference is on the ways OOM (out of memory errors) would happen on drivers vs executors. Hopefully, I'll be creating some content on OOM & other issues specifically and how to navigate through them. Hope that clarifies :)
Why to give such a large fraction (0.4) to User memory as in the end when the transformations will be performed in a particular stage , whether we give it a user defined function or any other function execution memory will be only used . So Whats exactly the role of User Memory ??
@@afaqueahmad7117 I am confused with user memory . As per spark documentation by default it is 40% of total memory. How can we check usage this memory in spark ui. Could you kindly please help to sort it. Kindly please make poc (proof of concept) video on resources usage by using GB's of data.
Hey @dileepn2479, as mentioned at 4.18, it's used for internal system level operations - these are not directly related to data processing but are essential for the proper functioning of the executor e.g. managing memory for JVM, networking during shuffling etc.. Hope this clarifies :)