Return-Path: Delivered-To: apmail-cassandra-user-archive@www.apache.org Received: (qmail 90548 invoked from network); 26 Jul 2010 16:26:55 -0000 Received: from unknown (HELO mail.apache.org) (140.211.11.3) by 140.211.11.9 with SMTP; 26 Jul 2010 16:26:55 -0000 Received: (qmail 62690 invoked by uid 500); 26 Jul 2010 16:26:53 -0000 Delivered-To: apmail-cassandra-user-archive@cassandra.apache.org Received: (qmail 62648 invoked by uid 500); 26 Jul 2010 16:26:52 -0000 Mailing-List: contact user-help@cassandra.apache.org; run by ezmlm Precedence: bulk List-Help: List-Unsubscribe: List-Post: List-Id: Reply-To: user@cassandra.apache.org Delivered-To: mailing list user@cassandra.apache.org Received: (qmail 62640 invoked by uid 99); 26 Jul 2010 16:26:52 -0000 Received: from nike.apache.org (HELO nike.apache.org) (192.87.106.230) by apache.org (qpsmtpd/0.29) with ESMTP; Mon, 26 Jul 2010 16:26:52 +0000 X-ASF-Spam-Status: No, hits=0.0 required=10.0 tests=FREEMAIL_FROM,RCVD_IN_DNSWL_NONE,SPF_PASS,T_TO_NO_BRKTS_FREEMAIL X-Spam-Check-By: apache.org Received-SPF: pass (nike.apache.org: domain of jshook@gmail.com designates 209.85.212.172 as permitted sender) Received: from [209.85.212.172] (HELO mail-px0-f172.google.com) (209.85.212.172) by apache.org (qpsmtpd/0.29) with ESMTP; Mon, 26 Jul 2010 16:26:46 +0000 Received: by pxi20 with SMTP id 20so147450pxi.31 for ; Mon, 26 Jul 2010 09:26:24 -0700 (PDT) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gmail.com; s=gamma; h=domainkey-signature:mime-version:received:received:in-reply-to :references:date:message-id:subject:from:to:content-type :content-transfer-encoding; bh=Wg0q5h3pg0opGQ7X5xW8kzCPRsYtgKpkmJWBbznVS80=; b=QBlNhBSfeviWtZMNQAGav9OkWevUrsD/Wp2XvXhYJ/bPJHZRiDHT7UAbz7zTKbTbxj WkYtIjqeItqRP34GP3BZBEIQfCurQ9EoTWMAgTeiileKwSAc1ai4fmRamqOYOk3cZYhY lBoBzlz9ErjtAZGTwNhgL72gk2tbgPCFDJZPw= DomainKey-Signature: a=rsa-sha1; c=nofws; d=gmail.com; s=gamma; h=mime-version:in-reply-to:references:date:message-id:subject:from:to :content-type:content-transfer-encoding; b=xFYEoHQGPmWyQfz6Fy+TOEzoN5UzmTatoB9C80ckBeybASpm+JOQ+4ggZOU8rB+llX Az/d9/7w3vESgHrxYJwr4bdoQ381IruMYxbwT96ZsIO9PYV7Qeb2YafDtGNhGgwwMmTI 6i8CRxeT322ggra7BxzgBBbl2KHy5A+x2lpUM= MIME-Version: 1.0 Received: by 10.114.92.10 with SMTP id p10mr11853927wab.214.1280161584294; Mon, 26 Jul 2010 09:26:24 -0700 (PDT) Received: by 10.114.176.9 with HTTP; Mon, 26 Jul 2010 09:26:24 -0700 (PDT) In-Reply-To: References: <79329e37-3f1e-4946-a807-81cdd86655f2@me.com> <795F2400-AB33-403A-A43B-D62DBEE59AB3@thelastpickle.com> Date: Mon, 26 Jul 2010 11:26:24 -0500 Message-ID: Subject: Re: Cassandra to store 1 billion small 64KB Blobs From: Jonathan Shook To: user@cassandra.apache.org Content-Type: text/plain; charset=ISO-8859-1 Content-Transfer-Encoding: quoted-printable X-Virus-Checked: Checked by ClamAV on apache.org Some possibilities open up when using OPP, especially with aggregate keys. This is more of an option when RF=3D=3Dcluster size, but not necessarily a good reason to make RF=3Dcluster size if you haven't already. For example, ':' and ';' make good boundary markers in aggregate keys, since they are already in ordinal order and not usually part of key names or hash values. If you structure your chunks like so: rowkey: : colname coldata ... then you can use the CF/row structure as a map. Asking for all rows between "customerABC:" and "customerABC;" will yield all files for abc. Asking for all rows between 'customerABC:file12' and 'customerABC;' would yield all files starting with the name 'file12', and so forth. One drawback is that you can't do searches on file name endings unless you iterate through the set of all files for that customer. This is where the secondary indexes may be better for describing your files. On the other hand, if you know the name of the file (via secondary index, ...) having a structure like this would potentially reduce your round-trips through Cassandra to get or store it. If you need another level of separation for row-length management, you can add another layer. A good 3rd layer would be the name/id of the first column in each row. I'd personally choose to store it in UTF-8 encoded hex. Your column names can simply be the first-byte offset of each chunk/column for reassembly or stream control/verification. 64MB per row, 1MB columns customerABC:file123:00000000 (colnames: 00000000, 00100000, 00200000, ...) customerABC:file123:04000000 (colnames: 04000000, 04100000, ... ) if 0xFFFFFFFF is not enough for the file size (4,294,967,295), then you can start with 10 or 12 digits instead (up to 2.8e+14) If you needed to add metadata to chunk groups/chunks, you can use column names which are disjoint from '0'-'F', as long as your API knows how to set your predicates up likewise. If there is at least one column name which is dependable in each chunk row, then you can use it as your predicate for "what's out there" queries. This avoids loading column data for the chunks when looking up names (row/file/... names). On the other hand, if you use an empty predicate, there is not an easy way to avoid tombstone rows unless you make another trip to Cassandra to verify. This is just more to consider. I'm not quite sure of the trade-offs here. I'd be curious of others' opinions on it. On Mon, Jul 26, 2010 at 10:06 AM, Michael Widmann wrote: > Okay . That really made a knot into my brain - It twist's a little bit no= w > I've to draw that=A0 on the whiteboard to understand it better ... but I'= ve > seen some very interesting cornerstones in your answer > for our project. > > really thanks a lot > mike > > 2010/7/26 aaron morton >> >> I see, got carried away thinking about it so here are some thoughts.... >> Your access patterns will determine the best storage design, so it's >> probably not the best solution. I would welcome thoughts from others. >> =3D> Standard CF: Chunks >> * key is chunk hash >> * col named 'data' col value is chunk data >> * another col with name "hash" and value of the hash. >> * could also store access data against the individual chunks here >> * probably cache lots of keys, and a only a few rows >> - to test if the hashes you have already exist do a multi get slice . Bu= t >> you need to specify a col to return the value for, and not the big data = one. >> So use the hash column. >> =3D> Standard CF: CustomerFiles >> * key is customer id >> * col per file name, col value perhaps latest version number or last >> accessed >> * cache keys and rows >> - to get all the files for a customer slice the customers row >> - if you want access data when you list all files for a customer, use a >> super CF with a super col for each file name. Store the meta data for th= e >> file in this CF *and* in the Files CF >> =3D> Super CF: Files >> * key is client_id.file_name >> * super column called "meta" with path / accessed etc. including version= s >> * super column called "current" with columns named "0001" and col values >> as the chunk hash >> * super column called "version.X" with the same format as above for >> current >> * cache keys and rows >> - assumes meta is shared across versions >> - to rebuild a file get_slice for all cols in the "current" super col an= d >> then do multi gets for the chunks >> - row grows as versions of file grows, but only storing links to the >> chunks so probably OK. Consider how many versions X how many chunks, the= n >> may way to make the number of rows grow instead of row size. Perhaps hav= e a >> FileVersions CF where the key includes the version number, then maintain >> information about the current version in both Files and FileVersions CFs= . >> Files CF would only ever have the current file, update access meta data = in >> both CF's. >> =3D> Standard CF: ChunkUsage >> * key is chunk hash >> * col name is a versioned file key cust_id.file_name.version col value i= s >> the count of usage in that version >> * no caching >> - cannot increase a counter until cassandra 0.7, so cannot keep a count = of >> chunk usage. >> - this is the reverse index to see where the chunk is used. >> - does not consider what is current, just all time usage. >> - not sure how much re-use their is, but row size grows with reuse. Shou= ld >> be ok for couple of million cols. >> >> Oh and if your going to use Hadoop / PIG to analyse the data in this >> beastie you need to think about that in the design. You'll probably want >> single CF to serve the queries with. >> Hope that helps >> Aaron >> >> On 26 Jul 2010, at 17:29, Michael Widmann wrote: >> >> Hi >> >> Wow that was lot of information... >> >> Think about users storing files online (means with their customer name) = - >> each customer maintains his own "hashtable" of files. Each File can cons= ist >> of some or several thousand entries (depends on the size of the whole fi= le). >> >> for example: >> >> File Test.doc=A0 consists of 3 * 64K Blobs=A0 - each blob does have the = hash >> value ABC - so we will only store one blob, one hash value and the entry >> that this blob is needed 3 times, so we try to avoid duplicate file data >> (and there's a lot of duplicates) >> >> Means our modell is: >> >> Filename: >> Test.doc >> Hashes: 3 >> Hash 1: ABC >> Hash 2: ABC >> Hash 3: ABC >> BLOB:ABC >> =A0Used: 3 >> =A0Binary: Data 1* >> >> Each customer does have: >> >> Customer:=A0 Customer:ID >> >> Filename: Test.doc >> =A0MetaData (Path / Accessed / modified / Size / compression / OS-Type f= rom >> / Security) >> =A0Version: 0 >> =A0Hash 1 / Hash 2 / Hash 3 >> >> Filename: Another.doc >> >> =A0 MetaData (Path / Accessed / modified / Size / compression / OS-Type = from >> / Security) >> =A0 Version: 0 >> =A0 Hash 1 / Hash 2 / Hash 3 >> =A0 Version: 1 >> =A0 Hash 1 / Hash 2 / Hash 4 >> =A0 Version: 2 >> =A0 Hash 3 / Hash 2 / Hash 2 >> >> Hope this clear some things :-) >> >> Mike >> >> >> 2010/7/26 Aaron Morton >>> >>> Some background reading.. >>> http://ria101.wordpress.com/2010/02/22/cassandra-randompartitioner-vs-o= rderpreservingpartitioner/ >>> >>> Not sure on your follow up question, so I'll just wildly blather on abo= ut >>> things :) >>> >>> My assumption of your data is you have 64K chunks that are identified b= y >>> a hash, which can somehow be grouped together into larger files (so the= re is >>> a "file name" of sorts). >>> >>> One possible storage design (assuming the Random Partitioner) is.... >>> >>> A Chunks CF, each row in this CF uses the hash of the chunk as it's key >>> and has is a single column with the chunk data. You could use more colu= mns >>> to store meta here. >>> >>> A ChunkIndex CF, each row uses the file name (from above) as the key an= d >>> has one column for each chunk in the file. The column name *could* be a= n >>> offset for the chunk and the column value could be the hash for the chu= nk. >>> Or you could use the chunk hash as the col name and the offset as the c= ol >>> value if needed. >>> >>> To rebuild the file read the entire row from the ChunkIndex, then make = a >>> series of multi gets to read all the chunks. Or you could lazy populate= the >>> ones you needed. >>> >>> This is all assuming that the 1000's comment below means you could want >>> to combine the chunks=A0 60+ MB chunks. It would be easier to keep all = the >>> chunks together in one row, if you are going to have large (unbounded) = file >>> size this may not be appropriate. >>> >>> You could also think about using the order preserving partitioner, and >>> using a compound key for each row such as "file_name_hash.offset" . The= n by >>> using the get_range_slices to scan the range of chunks for a file you w= ould >>> not need to maintain a secondary index. Some drawbacks to that approach= , >>> read the article above. >>> >>> Hope the helps >>> Aaron >>> >>> >>> On 26 Jul, 2010,at 04:01 PM, Michael Widmann >>> wrote: >>> >>> Thanks for this detailed description ... >>> >>> You mentioned the secondary index in a standard column, would it be >>> better to build several indizes? >>> Is that even possible to build a index on for example 32 columns? >>> >>> The hint with the smaller boxes is very valuable! >>> >>> Mike >>> >>> 2010/7/26 Aaron Morton >>>> >>>> For what it's worth... >>>> >>>> * Many smaller boxes with local disk storage are preferable to 2 with >>>> huge NAS storage. >>>> * To cache the hash values look at the KeysCached setting in the >>>> storage-config >>>> * There are some row size limits see >>>> http://wiki.apache.org/cassandra/CassandraLimitations >>>> * If you wanted to get 1000 blobs, rather then group them in a single >>>> row using a super column consider building a secondary index in a stan= dard >>>> column. One CF for the blobs using your hash, one CF that uses whateve= r they >>>> grouping key is with a col for every blobs hash value. Read from the i= ndex >>>> first, then from the blobs themselves. >>>> >>>> Aaron >>>> >>>> >>>> On 24 Jul, 2010,at 06:51 PM, Michael Widmann >>>> wrote: >>>> >>>> Hi Jonathan >>>> >>>> Thanks for your very valuable input on this. >>>> >>>> I maybe didn't enough explanation - so I'll try to clarify >>>> >>>> Here are some thoughts: >>>> >>>> binary data will not be indexed - only stored. >>>> The file name to the binary data (a hash) should be indexed for search >>>> We could group the hashes in 62 "entry" points for search retrieving -= > >>>> i think suprcolumns (If I'm right in terms) (a-z,A_Z,0-9) >>>> the 64k Blobs meta data (which one belong to which file) should be >>>> stored separate in cassandra >>>> For Hardware we rely on solaris / opensolaris with ZFS in the backend >>>> Write operations occur much more often than reads >>>> Memory should hold the hash values mainly for fast search (not the >>>> binary data) >>>> Read Operations (restore from cassandra) may be async - (get about 100= 0 >>>> Blobs) - group them restore >>>> >>>> So my question is too: >>>> >>>> 2 or 3 Big boxes or 10 till 20 small boxes for storage... >>>> Could we separate "caching" - hash values CFs cashed and indexed - >>>> binary data CFs not ... >>>> Writes happens around the clock - on not that tremor speed but >>>> constantly >>>> Would compaction of the database need really much disk space >>>> Is it reliable on this size (more my fear) >>>> >>>> thx for thinking and answers... >>>> >>>> greetings >>>> >>>> Mike >>>> >>>> 2010/7/23 Jonathan Shook >>>>> >>>>> There are two scaling factors to consider here. In general the worst >>>>> case growth of operations in Cassandra is kept near to O(log2(N)). An= y >>>>> worse growth would be considered a design problem, or at least a high >>>>> priority target for improvement. =A0This is important for considering >>>>> the load generated by very large column families, as binary search is >>>>> used when the bloom filter doesn't exclude rows from a query. >>>>> O(log2(N)) is basically the best achievable growth for this type of >>>>> data, but the bloom filter improves on it in some cases by paying a >>>>> lower cost every time. >>>>> >>>>> The other factor to be aware of is the reduction of binary search >>>>> performance for datasets which can put disk seek times into high >>>>> ranges. This is mostly a direct consideration for those installations >>>>> which will be doing lots of cold reads (not cached data) against larg= e >>>>> sets. Disk seek times are much more limited (low) for adjacent or nea= r >>>>> tracks, and generally much higher when tracks are sufficiently far >>>>> apart (as in a very large data set). This can compound with other >>>>> factors when session times are longer, but that is to be expected wit= h >>>>> any system. Your storage system may have completely different >>>>> characteristics depending on caching, etc. >>>>> >>>>> The read performance is still quite high relative to other systems fo= r >>>>> a similar data set size, but the drop-off in performance may be much >>>>> worse than expected if you are wanting it to be linear. Again, this i= s >>>>> not unique to Cassandra. It's just an important consideration when >>>>> dealing with extremely large sets of data, when memory is not likely >>>>> to be able to hold enough hot data for the specific application. >>>>> >>>>> As always, the real questions have lots more to do with your specific >>>>> access patterns, storage system, etc I would look at the benchmarking >>>>> info available on the lists as a good starting point. >>>>> >>>>> >>>>> On Fri, Jul 23, 2010 at 11:51 AM, Michael Widmann >>>>> wrote: >>>>> > Hi >>>>> > >>>>> > We plan to use cassandra as a data storage on at least 2 nodes with >>>>> > RF=3D2 >>>>> > for about 1 billion small files. >>>>> > We do have about 48TB discspace behind for each node. >>>>> > >>>>> > now my question is - is this possible with cassandra - reliable - >>>>> > means >>>>> > (every blob is stored on 2 jbods).. >>>>> > >>>>> > we may grow up to nearly 40TB or more on cassandra "storage" data .= .. >>>>> > >>>>> > anyone out did something similar? >>>>> > >>>>> > for retrieval of the blobs we are going to index them with an >>>>> > hashvalue >>>>> > (means hashes are used to store the blob) ... >>>>> > so we can search fast for the entry in the database and combine the >>>>> > blobs to >>>>> > a normal file again ... >>>>> > >>>>> > thanks for answer >>>>> > >>>>> > michael >>>>> > >>>> >>>> >>>> >>>> -- >>>> bayoda.com - Professional Online Backup Solutions for Small and Medium >>>> Sized Companies >>> >>> >>> >>> -- >>> bayoda.com - Professional Online Backup Solutions for Small and Medium >>> Sized Companies >> >> >> >> -- >> bayoda.com - Professional Online Backup Solutions for Small and Medium >> Sized Companies >> > > > > -- > bayoda.com - Professional Online Backup Solutions for Small and Medium Si= zed > Companies >