New version page

Berkeley COMPSCI 186 - Lecture Final

Documents in this Course
Load more
Upgrade to remove ads
Upgrade to remove ads
Unformatted text preview:

AdministriviaOffice HoursAs you study...Database Lessons to Live ByRecall Lecture 1!!Simplicity is BeautifulBulk Processing & I/O Go TogetherOptimize the Memory HierarchyQuery Processing is PredictableApplied Algorithm AnalysisIndexing Is Simple, PowerfulNot enough memory? Partition!Declarative languages are great!SQL: The good, the bad, the uglyQuery Operators & OptimizationDatabase DesignCC & Recovery: House SpecialtiesThe Rebirth of Information RetrievalDatabases: The natural way to leverage parallelism & distribution“More, more, I’m still not satisfied” -- Tom LehrerBut wait, there’s more!Parting ThoughtsAdministrivia•Final Exam–Tuesday, 5/20, 5-8 pm–Cumulative, stress end of semester–2 cribsheets•Final Review Session–Watch for announcementOffice Hours•Next week–Tentative office hours on 5/15, watch web pageAs you study...•"Reading maketh a full man; conference a ready man; and writing an exact man." -Francis Bacon •"If you want truly to understand something, try to change it." -Kurt Lewin •"I hear and I forget. I see and I remember. I do and I understand." -Chinese Proverb. •"Knowledge is a process of piling up facts; wisdom lies in their simplification." -Martin H. FischerDatabase Lessons to Live By“If we do well here, we shall do well there: I can tell you no more if I preach a whole year”-- John Edwin (1749-1790)Recall Lecture 1!!•Lessons of Data Independence–High-level, declarative programming–Maintenance in the face of change•Automatic re-optimization•Data integrity–Declarative consistency (constraints, FDs)–Concurrent access, recovery from crashes.Simplicity is Beautiful•The relational model is simple–simple query language means simple implementation model•basically just indexes, join algorithms, sorting, grouping!–simple data model means easy schema evolution–simple data model provides clean analysis of schemas (FD’s & NF’s are essentially automatic)–Every other structured data model has proved to be a wash•XML has found a niche, but not as a database•There’s a reason that the backend of web search looks so much like a relational database.Bulk Processing & I/O Go Together•Disks provide data a page at a time•Databases deal with data a set at a time–sets usually bigger than a page–means I/O costs are usually justified.–much better than other techniques, which are “object-at-a-time”•Set-at-a-time allows for optimization–can do bulk operations (e.g. sort or hash) –or can do things tuple-at-a-time (e.g. nested loops)Optimize the Memory Hierarchy•DBMS worries about Disk vs. RAM–spend lotsa CPU cycles planning disk access–I/O cost “hides” the think time•Similar hierarchies exist in other parts of a computer –various caches on and off CPU chips–less time to spare optimizing here•Change is happening here!–Disk is the new tape–Flash is the new disk–RAM is really bigQuery Processing is Predictable•Big queries take many predictable steps–unlike typical OS workloads, which depend on what small task users decide to do next•DBMSs can use this knowledge to optimize–For caching, prefetching, admission control, memory allocation, etc.•These lessons should be applied whenever you know your access patterns–again, especially for bulk operations!Applied Algorithm Analysis•Know the practical costs of your algorithms–The optimizer needs to know anyway–How many disk I/O’s really needed to access a B+Tree?•In many applications, the bottlenecks determine the cost model–e.g. I/O is traditional DB bottleneck–in another setting it might be network, or processor cache locality–this affects the practical analysis of the algorithmIndexing Is Simple, Powerful•Hash indexes easy and quick for equality–worth reading about linear hashing in the text•Trees can be used for just about anything else! –each tree level partitions the dataset–labels in the tree “direct query traffic” to the right data–“all” you need to think about in designing a tree is how to partition, and how to label!Not enough memory? Partition!•Traditional main-memory algorithms can be extended to disk-based algorithms–partition input (runs for sorting, partitions for hash-table)–process partitions (sort runs, hash partitions)–merge partitions (merge runs, concatenate partitions)•Sorting & hashing very similar!–their I/O patterns are “dual”Declarative languages are great!•Simple: say what you want, not how to get it!•Should correctly convert to an imperative language–Codd’s Theorem says rel. calc. = rel. alg.–no such theorem for text ranking :-(•If you can convert in different ways, you get to optimize!–hides complexity from user–accomodates changes in database without requiring applications to be recompiled.•Especially important when–App Rate of Change << Physical Rate of Change•A reborn trend in computing–Declarative networking, security, robotics, natural language processing, distributed systems, …SQL: The good, the bad, the ugly•SQL is very simple–SELECT..FROM..WHERE•Well...SQL is kind of tricky–aggregation, GROUP BY, HAVING•OK, OK. SQL is complicated!–duplicates & NULLs–Subqueries–dups/NULLs/subqueries/aggregation together!•Remember: SQL is not entirely declarative!!!•But, it beats the heck out of writing (and maintaining!) C++ or Java programs for every queryQuery Operators & Optimization•Query operators are actually all similar:–Sorting, Hashing, Iteration•Query Optimization: 3-part harmony–define a plan space–estimate costs for plans–algorithm to search in the plan space for cheapest•Research on each of the 3 pieces goes on independently! (Usually…)•Nice clean model for attacking a hard problemDatabase Design•(And you thought SQL was confusing!)•This is not simple stuff!!–requires a lot of thought, a lot of tools–there’s no cookbook to follow–decisions can make a hu ge difference down the road!•The basic steps we studied (conceptual design, schema refinement, physical design) break up the problem somewhat, but also interact with each other•Complexity in DB design pays off at query time, and in consistency–vs. filesCC & Recovery: House Specialties•RDBMSs nailed concurrency and reliability–transactions & 2-phase locking–write-ahead-logging–details are tricky, worked out over 20 years!•Also models for relaxing transactions–Lower degrees of consistency•Other

View Full Document
Download Lecture Final
Our administrator received your request to download this document. We will send you the file to your email shortly.
Loading Unlocking...

Join to view Lecture Final and access 3M+ class-specific study document.

We will never post anything without your permission.
Don't have an account?
Sign Up

Join to view Lecture Final 2 2 and access 3M+ class-specific study document.


By creating an account you agree to our Privacy Policy and Terms Of Use

Already a member?