I think this statement is not true: "By distributing by (and preferably ordering by) user_id, we can minimize seek time in the table because Hive knows where all entries pertaining to a specific user are stored." I think it is not true whether the table is bucketed on user_id or not (assuming that user_id is not a partition column or indexed column).
-----Original Message----- From: Mark Grover [mailto:mgro...@oanda.com] Sent: Tuesday, September 06, 2011 2:36 PM To: user@hive.apache.org Cc: Travis Powell; Baiju Devani; Bob Tiernay Subject: Re: Best practices for storing data on Hive Thanks for your reply, Travis. I was under the impression that for Hive to make use of sorted structure of data (i.e. for the table named "data" in your example), the metadata of the table (specified during table creation) has to advertise such property. However, I don't see any special metadata specifying such property when "data" table was created. Is that true? If so, is such metadata specified by using CLUSTERED BY and SORTED BY clauses during table creation? On 11-09-06 03:50 PM, Travis Powell wrote: > Hi Mark, > > When we load data into Hive, we use a staging table to dynamically partition > our data. This might help you too. > > We create our initial table and our staging table: > > DROP TABLE IF EXISTS staging_data; > CREATE TABLE staging_data ( ... ) > ROW FORMAT DELIMITED FIELDS TERMINATED BY ','; > CREATE TABLE data ( ... ) > PARTITIONED BY (dt STRING, hour, INT) > ROW FORMAT DELIMITED FIELDS TERMINATED BY ',' STORED AS SEQUENCEFILE; > > INSERT OVERWRITE TABLE data PARTITION(dt, hour) SELECT q.*, > to_date(q.session_timestamp) AS dt, hour(q.session_timestamp) AS hour FROM > staging_session q ORDER BY user_id DISTRIBUTE BY user_id; > > > So..... > By distributing by (and preferably ordering by) user_id, we can minimize seek > time in the table because Hive knows where all entries pertaining to a > specific user are stored. Partitions by time have the best performance, > because chances are almost every query will have some time-related component > in it (and it spreads out the data among partitions fairly well.) > > Let me know if this works for you. We start every job with those first few > lines of Hive script. It works well for us. > > Thanks, > > Travis Powell > > -----Original Message----- > From: Mark Grover [mailto:mgro...@oanda.com] > Sent: Tuesday, September 06, 2011 12:39 PM > To: user@hive.apache.org > Cc: wd; Bob Tiernay; Baiju Devani > Subject: Re: Best practices for storing data on Hive > > Thanks for the response, wd. > > I would REALLY APPRECIATE if other people can share their views as well. > > Here are the possible solutions that I have thought about to the problem > (see original email for description of problem): > > 1) Multiple partitions: We would partition the table by day and userId. > However, given the amount of users that visit our website (hundreds of > thousands of unique users every day), this would lead to a large number > of partitions (and rather small file sizes, ranging from a couple of > bytes to a couple of KB). From the documentation I've read online, it > seems that Hive/Hadoop weren't designed to deal with such small file > sizes and such a situation should be avoided if possible. > We had a scenario previously where we were partitioning by day and hour > and because of the sheer number of partitions queries like "select * > from<table> LIMIT 1;" were taking very long and even failed because of > "Java out of Heap space" errors. My guess is that the master node was > munching through all these partitions and couldn't deal with the large > number of partitions. > > 2) Use of data locality: We could keep the data partitioned by day and > bucketed by userId. Within each bucket sort the data by the (userId, > time). This way we could keep the data related to each userId together > within a daily partition and if Hive could be made aware of this sorting > order and could make use of this order to improve search/query times, > that would alleviate the problem quite a bit. The big question here is: > Does Hive leverage sorting order of data within a partition bucket when > running (most/all?) queries, where possible? > > > 3) Using an index: As wd mentioned, Hive 0.7 introduces the notion on an > index. If I do index on userId, given that we can hundreds of thousands > of unique users per day, would indexing prove to be a good move? Are > there people who are using it for similar purposes or on a similar scale? > > > 4) Using 2 "orthogonal tables": As mentioned in my original email (see > below), we could have 2 independent tables, one which stores data > partitioned by day and other partitioned by userId. For maintaining > partitions in userId partitioned table, I am planning to do the following: > In the nightly job, if userId=X visited the website previous day, we > create a partition for userId=X if it doesn't already exist. Once the > partition is created, all clicks for that user Id on the day for in > question are put in a single file and dropped in the userId=X folder on > HDFS. This method could be used to simulate an "append" to the Hive > table. The file would only be a few bytes to a few KB and the format of > the table would be sequence file. > > What are your thoughts about the above 4 methods? Any particular likes > or dislikes? Any comments, suggestions would be helpful. > > Thank you again in advance! > > Mark > > On 11-09-04 04:01 AM, wd wrote: >> Hive support more than one partitions, have your tried? Maybe you can >> create to partitions named as date and user. >> >> Hive 0.7 also support index, maybe you can have a try. >> >> On Sat, Sep 3, 2011 at 1:18 AM, Mark Grover<mgro...@oanda.com> wrote: >>> Hello folks, >>> I am fairly new to Hive and am wondering if you could share some of the >>> best practices for storing/querying data with Hive. >>> >>> Here is an example of the problem I am trying to solve. >>> >>> The traffic to our website is logged in files that contain information >>> about clicks from various users. >>> Simplified, the log file looks like: >>> t_1, ip_1, userid_1 >>> t_2, ip_2, userid_2 >>> t_3, ip_3, userid_3 >>> ... >>> >>> where t_i represents time of the click, ip_i represents ip address where >>> the click originated from, and userid_i represents the user ID of the user. >>> >>> Since the clicks are logged on an ongoing basis, partitioning our Hive >>> table by day seemed like the obvious choice. Every night we upload the data >>> from the previous day into a new partition. >>> >>> However, we would also want the capability to find all log lines >>> corresponding to a particular user. With our present partitioning scheme, >>> all day partitions are searched for that user ID but this takes a long >>> time. I am looking for ideas/suggestions/thoughts/comments on how to reduce >>> this time. >>> >>> As a solution, I am thinking that perhaps we could have 2 independent >>> tables, one which stores data partitioned by day and the other partitioned >>> by userId. With the second table partitioned by userId, I will have to find >>> some way of maintaining the partitions since Hive doesn't support appending >>> of files. Also, this seems suboptimal, since we are doubling that the >>> amount of data that we store. What do you folks think of this idea? >>> >>> Do you have any other suggestions on how we can approach this problem? >>> >>> What have other people in similar situations done? Please share. >>> >>> Thank you in advance! >>> Mark >>> -- Mark Grover, Business Intelligence Analyst OANDA Corporation www: oanda.com www: fxtrade.com e: mgro...@oanda.com "Best Trading Platform" - World Finance's Forex Awards 2009. "The One to Watch" - Treasury Today's Adam Smith Awards 2009.
