Introduction to Information Retrieval Introduction to Information Retrieval Lecture 4: Index Construction Related to Chapter 4:

Introduction to Information Retrieval This lecture  Scaling index construction  Distributed indexing  Dynamic indexing 2

Introduction to Information Retrieval  Documents are parsed to extract words and these are saved with the Document ID. I did enact Julius Caesar I was killed i' the Capitol; Brutus killed me. Doc 1 So let it be with Caesar. The noble Brutus hath told you Caesar was ambitious Doc 2 Inverted index construction Sec. 4.2

Introduction to Information Retrieval Key step  After all documents have been parsed, the inverted file is sorted by terms and then by Doc# We focus on this sort step. We have 100M items to sort. Sec. 4.2

Introduction to Information Retrieval Scaling index construction Sec. 4.2

Introduction to Information Retrieval Hardware basics  Many design decisions in information retrieval are based on the characteristics of hardware  We begin by reviewing hardware basics Sec. 4.1

Introduction to Information Retrieval Hardware basics  Access to data in memory is much faster than access to data on disk.  Servers used in IR systems now typically have several GB of main memory, sometimes tens of GB.  Available disk space is several orders of magnitude larger. Sec. 4.1

Introduction to Information Retrieval Hardware basics Sec. 4.1

Introduction to Information Retrieval Hardware basics 9

Introduction to Information Retrieval RCV1: Our collection for this lecture  Shakespeare’s collected works definitely aren’t large enough for demonstrating many of the points in this course.  As an example for applying scalable index construction algorithms, we will use the Reuters RCV1 collection.  This is one year of Reuters newswire.  It isn’t really large enough either, but it’s publicly available and is at least a more plausible example. Sec. 4.2

Introduction to Information Retrieval A Reuters RCV1 document Sec. 4.2

Introduction to Information Retrieval Reuters RCV1 statistics  symbolstatistic value  N documents 800,000  Mterms400,000  tokens100,000,000 Sec. 4.2

Introduction to Information Retrieval Sort-based index construction  As we build the index, we parse docs one at a time.  At 12 bytes (4+4+4) per token entry (term, doc, freq), demands a lot of space for large collections.  Why 4 bits per term?  “Term to term-id” dictionary and “term-id to doc-id” dictionary.  T = 100,000,000 in the case of RCV1  We can do this in memory in 2009, but typical collections are much larger. E.g., the New York Times provides an index of >150 years of newswire  Thus: We need to store intermediate results on disk. Sec. 4.2

Introduction to Information Retrieval Sort using disk as “memory”?  Can we use the same index construction algorithm for larger collections, but by using disk instead of memory?  Sorting T = 100,000,000 records on disk is too slow  Too many (random) disk seeks. Sec. 4.2

Introduction to Information Retrieval The solution?  We need to use an external sorting algorithm, that is, one that uses disk.  For acceptable speed, the central requirement of such an algorithm is that it minimize the number of random disk seeks during sorting.  Sequential disk reads are far faster than seeks.  Two methods:  BSBI  SPIMI (not present in our course) 15

Introduction to Information Retrieval BSBI: Blocked Sort-Based Indexing  (i) segments the collection into parts of equal size,  (ii) sort and invert the termID–docID pairs of each part in memory,  First, we sort the termID–docID pairs.  Next, we collect all termID–docID pairs with the same termID into a postings list  (iii) stores intermediate sorted results on disk,  (iv) merges all intermediate results into the final index. 16

Introduction to Information Retrieval Sec. 4.2

Introduction to Information Retrieval Example  Must now sort 100M records.  Define a Block ~ 10M such records  Can easily fit a couple into memory.  Will have 10 such blocks to start with. Sec. 4.2

Introduction to Information Retrieval How to merge the sorted runs? Sec. 4.2

Introduction to Information Retrieval How to merge the sorted runs?  Open all block files simultaneously.  Maintain small read buffers for the ten blocks and a write buffer for the final merged index we are writing.  In each iteration, we select the lowest termID that has not been processed yet.  All postings lists for this termID are read and merged, and the merged list is written back to disk.  Each read buffer is refilled from its file when necessary. 20

Introduction to Information Retrieval How to merge the sorted runs?  Providing you read decent-sized chunks of each block into memory and then write out a decent-sized output chunk, then you’re not killed by disk seeks Sec. 4.2

Introduction to Information Retrieval Remaining problem with sort-based algorithm  Our assumption was: we can keep the dictionary in memory.  We need the dictionary (which grows dynamically) in order to implement a term to termID mapping.  Actually, we could work with term,docID postings instead of termID,docID postings  But then intermediate files become very large. (We would end up with a scalable, but very slow index construction method)  The solution: SPMI method. Sec. 4.3

Introduction to Information Retrieval This lecture  Scaling index construction  Distributed indexing  Dynamic indexing 23

Introduction to Information Retrieval An example of distributed processing 24

Introduction to Information Retrieval Distributed indexing  For web-scale indexing (don’t try this at home!):  must use a distributed computing cluster  a group of tightly coupled computers that work together closely  Individual machines are fault-prone  Can unpredictably slow down or fail  How do we exploit such a pool of machines?  Estimate: Google ~1 million servers, 3 million processors/cores (Gartner 2007) Sec. 4.4

Introduction to Information Retrieval Distributed indexing  Maintain a master machine directing the indexing job  considered “safe”.  Break up indexing into sets of (parallel) tasks.  Master machine assigns each task to an idle machine from a pool. Sec. 4.4

Introduction to Information Retrieval Parallel tasks Sec. 4.4

Introduction to Information Retrieval Parsers Sec. 4.4

Introduction to Information Retrieval Inverters Sec. 4.4

Introduction to Information Retrieval Data flow splits Parser Master a-fg-pq-z a-fg-pq-z a-fg-pq-z Inverter Postings a-f g-p q-z assign Map phase Segment files Reduce phase Sec. 4.4

Introduction to Information Retrieval MapReduce Sec. 4.4

Introduction to Information Retrieval This lecture  Scaling index construction  Distributed indexing  Dynamic indexing 32

Introduction to Information Retrieval Dynamic indexing  Up to now, we have assumed that collections are static.  They rarely are:  Documents come in over time and need to be inserted.  Documents are deleted and modified.  This means that the dictionary and postings lists have to be modified:  Postings updates for terms already in dictionary  New terms added to dictionary Sec. 4.5

Introduction to Information Retrieval The simplest approach  Periodically reconstruct the index.  This is a good solution if  The number of changes over time is small  Delay in making new documents searchable is acceptable  Enough resources are available to construct a new index while the old one is still available for querying. 34

Introduction to Information Retrieval Another approach Sec. 4.5

Introduction to Information Retrieval Merging 36

Introduction to Information Retrieval Logarithmic Merging 37 Main index K: the number of times, auxiliary is filled.

Introduction to Information Retrieval Sec. 4.5

Introduction to Information Retrieval Logarithmic merge Sec. 4.5 

Introduction to Information Retrieval Final considerations  So far, postings lists are ordered with respect to docID.  As we see in Chapter 5, this is advantageous for compression.  However, this structure for the index is not optimal when we build ranked retrieval systems (chapter 6 and 7).  In ranked retrieval, postings are often ordered according to weight or impact.  In a docID-sorted index, new documents are always inserted at the end of postings lists.  In an impact-sorted index, the insertion can occur anywhere, thus complicating the update of the inverted index. 40