Re: Achieving isolation on single row modifications with batch_mutate
I'm chunking up a larger blob. Basically the size of each row can vary (averages around 500K - 1MB), with some outliers in the 50 MB range. However, when I do an update, I can usually just read/update a portion of that blob. A lot of my read operations can also work on a smaller chunk. The number of columns is going to depend on the size of the blob itself. I'm also considering using supercolumns to have higher save granularity. My biggest problem is that I will have to update these rows a lot (several times a day) and often very quickly (process 15 thousand in 2-3 minutes). While I think I could probably scale up with a lot of hardware to meet that load, it seems like I'm doing much much more work than I need to (processing 15 GB of data in 2-3 minutes as opposes to 100 MB). I also worry about handling our future data size needs. I can split the blob up without a lot of extra complexity but am worried about how to have readers read a non-corrupted version of the object, since sometimes I'll have to update multiple chunks as one unit. From: Tyler Hobbs ty...@riptano.com To: user@cassandra.apache.org Sent: Tue, November 30, 2010 12:57:07 AM Subject: Re: Achieving isolation on single row modifications with batch_mutate In this case, it sounds like you should combine columns A and B if you are writing them both at the same time, reading them both at the same time, and need them to be consistent. Obviously, you're probably dealing with more than two columns here, but there's generally not any value in splitting something into multiple columns if you're always writing and reading all of them at the same time. Or are you talking about chunking huge blobs across a row? - Tyler On Sat, Nov 27, 2010 at 10:12 AM, E S tr1skl...@yahoo.com wrote: I'm trying to figure out the best way to achieve single row modification isolation for readers. As an example, I have 2 rows (1,2) with 2 columns (a,b). If I modify both rows, I don't care if the user sees the write operations completed on 1 and not on 2 for a short time period (seconds). I also don't care if when reading row 1 the user gets the new value, and then on a re-read gets the old value (within a few seconds). Because of this, I have been planning on using a consistency level of one. However, if I modify both columns A,B on a single row, I need both changes on the row to be visible/invisible atomically. It doesn't matter if they both become visible and then both invisible as the data propagates across nodes, but a half-completed state on an initial read will basically be returning corrupt data given my apps consistency requirements. My understanding from the FAQ that this single row multicolumn change provides no read isolation, so I will have this problem. Is this correct? If so: Question 1: Is there a way to get this type of isolation without using a distributed locking mechanism like cages? Question 2: Are there any plans to implement this type of isolation within Cassandra? Question 3: If I went with a distributed locking mechanism, what consistency level would I need to use with Cassandra? Could I still get away with a consistency level of one? It seems that if the initial write is done in a non-isolated way, but if cross-node row synchronizations are done all or nothing, I could still use one. Question 4: Does anyone know of a good c# alternative to cages/zookeeper? Thanks for any help with this!
Re: Achieving isolation on single row modifications with batch_mutate
On Sat, Nov 27, 2010 at 10:12 AM, E S tr1skl...@yahoo.com wrote: I'm trying to figure out the best way to achieve single row modification isolation for readers. I have a lot of No's for you. :) As an example, I have 2 rows (1,2) with 2 columns (a,b). If I modify both rows, I don't care if the user sees the write operations completed on 1 and not on 2 for a short time period (seconds). I also don't care if when reading row 1 the user gets the new value, and then on a re-read gets the old value (within a few seconds). Because of this, I have been planning on using a consistency level of one. However, if I modify both columns A,B on a single row, I need both changes on the row to be visible/invisible atomically. It doesn't matter if they both become visible and then both invisible as the data propagates across nodes, but a half-completed state on an initial read will basically be returning corrupt data given my apps consistency requirements. My understanding from the FAQ that this single row multicolumn change provides no read isolation, so I will have this problem. Is this correct? If so: Question 1: Is there a way to get this type of isolation without using a distributed locking mechanism like cages? No. Question 2: Are there any plans to implement this type of isolation within Cassandra? No. Question 3: If I went with a distributed locking mechanism, what consistency level would I need to use with Cassandra? Could I still get away with a consistency level of one? Maybe. If you want to guarantee that you see the most recent write, then ONE will not be high enough. But if all you care about is seeing all of the update or none of it, then ONE + locking will be fine. Question 4: Does anyone know of a good c# alternative to cages/zookeeper? No. -- Jonathan Ellis Project Chair, Apache Cassandra co-founder of Riptano, the source for professional Cassandra support http://riptano.com
Re: Achieving isolation on single row modifications with batch_mutate
It's hard to tell without knowing the the nature of the data you're writing, but you might want to think about whether you can embed any sort of version number and/or checksum into the column names of the chunk columns. That way, you could very easily determine that the data you wanted to retrieve was not yet available for reading. Are you able do your partial blob updates on an entire chunk at a time or do you need to read the blob chunk, modify a portion of it, and then write it back? If it's the former, then it might be possible for this to be accomplished without a locking solution. Ed On Sat, Nov 27, 2010 at 8:12 AM, E S tr1skl...@yahoo.com wrote: I'm trying to figure out the best way to achieve single row modification isolation for readers. As an example, I have 2 rows (1,2) with 2 columns (a,b). If I modify both rows, I don't care if the user sees the write operations completed on 1 and not on 2 for a short time period (seconds). I also don't care if when reading row 1 the user gets the new value, and then on a re-read gets the old value (within a few seconds). Because of this, I have been planning on using a consistency level of one. However, if I modify both columns A,B on a single row, I need both changes on the row to be visible/invisible atomically. It doesn't matter if they both become visible and then both invisible as the data propagates across nodes, but a half-completed state on an initial read will basically be returning corrupt data given my apps consistency requirements. My understanding from the FAQ that this single row multicolumn change provides no read isolation, so I will have this problem. Is this correct? If so: Question 1: Is there a way to get this type of isolation without using a distributed locking mechanism like cages? Question 2: Are there any plans to implement this type of isolation within Cassandra? Question 3: If I went with a distributed locking mechanism, what consistency level would I need to use with Cassandra? Could I still get away with a consistency level of one? It seems that if the initial write is done in a non-isolated way, but if cross-node row synchronizations are done all or nothing, I could still use one. Question 4: Does anyone know of a good c# alternative to cages/zookeeper? Thanks for any help with this!
Re: Achieving isolation on single row modifications with batch_mutate
I'm a little confused about #3. Hopefully this clarifying question won't turn the one maybe into a no :). I'm fine not reading the latest data, as long as on each individual read I see all or none of the operations that occurred for a single one row batch_mutate. My concern is do I have to lock the reads until they have propagated to all nodes. If I do a batch_mutate with a consistency of ONE onto one row, during the write operation to the one node a reader can see partial changes. Once the batch mutate completes, the change has not been propagated to the other nodes. On a per row basis, are the changes to other nodes pushed in an isolated manner? If not, it seems like I would have to write with a consistency of ALL and lock around that. - Original Message From: Jonathan Ellis jbel...@gmail.com To: user user@cassandra.apache.org Sent: Tue, November 30, 2010 9:50:51 AM Subject: Re: Achieving isolation on single row modifications with batch_mutate On Sat, Nov 27, 2010 at 10:12 AM, E S tr1skl...@yahoo.com wrote: I'm trying to figure out the best way to achieve single row modification isolation for readers. I have a lot of No's for you. :) As an example, I have 2 rows (1,2) with 2 columns (a,b). If I modify both rows, I don't care if the user sees the write operations completed on 1 and not on 2 for a short time period (seconds). I also don't care if when reading row 1 the user gets the new value, and then on a re-read gets the old value (within a few seconds). Because of this, I have been planning on using a consistency level of one. However, if I modify both columns A,B on a single row, I need both changes on the row to be visible/invisible atomically. It doesn't matter if they both become visible and then both invisible as the data propagates across nodes, but a half-completed state on an initial read will basically be returning corrupt data given my apps consistency requirements. My understanding from the FAQ that this single row multicolumn change provides no read isolation, so I will have this problem. Is this correct? If so: Question 1: Is there a way to get this type of isolation without using a distributed locking mechanism like cages? No. Question 2: Are there any plans to implement this type of isolation within Cassandra? No. Question 3: If I went with a distributed locking mechanism, what consistency level would I need to use with Cassandra? Could I still get away with a consistency level of one? Maybe. If you want to guarantee that you see the most recent write, then ONE will not be high enough. But if all you care about is seeing all of the update or none of it, then ONE + locking will be fine. Question 4: Does anyone know of a good c# alternative to cages/zookeeper? No. -- Jonathan Ellis Project Chair, Apache Cassandra co-founder of Riptano, the source for professional Cassandra support http://riptano.com
Re: Achieving isolation on single row modifications with batch_mutate
In this case, it sounds like you should combine columns A and B if you are writing them both at the same time, reading them both at the same time, and need them to be consistent. Obviously, you're probably dealing with more than two columns here, but there's generally not any value in splitting something into multiple columns if you're always writing and reading all of them at the same time. Or are you talking about chunking huge blobs across a row? - Tyler On Sat, Nov 27, 2010 at 10:12 AM, E S tr1skl...@yahoo.com wrote: I'm trying to figure out the best way to achieve single row modification isolation for readers. As an example, I have 2 rows (1,2) with 2 columns (a,b). If I modify both rows, I don't care if the user sees the write operations completed on 1 and not on 2 for a short time period (seconds). I also don't care if when reading row 1 the user gets the new value, and then on a re-read gets the old value (within a few seconds). Because of this, I have been planning on using a consistency level of one. However, if I modify both columns A,B on a single row, I need both changes on the row to be visible/invisible atomically. It doesn't matter if they both become visible and then both invisible as the data propagates across nodes, but a half-completed state on an initial read will basically be returning corrupt data given my apps consistency requirements. My understanding from the FAQ that this single row multicolumn change provides no read isolation, so I will have this problem. Is this correct? If so: Question 1: Is there a way to get this type of isolation without using a distributed locking mechanism like cages? Question 2: Are there any plans to implement this type of isolation within Cassandra? Question 3: If I went with a distributed locking mechanism, what consistency level would I need to use with Cassandra? Could I still get away with a consistency level of one? It seems that if the initial write is done in a non-isolated way, but if cross-node row synchronizations are done all or nothing, I could still use one. Question 4: Does anyone know of a good c# alternative to cages/zookeeper? Thanks for any help with this!