Fast HTAP over loosely coupled Nodes

Size: px

Start display at page:

Download "Fast HTAP over loosely coupled Nodes"

Clemence Craig
5 years ago
Views:

1 Wildfire: Fast HTAP over loosely coupled Nodes R. Barber, C. Garcia-Arellano, R. Grosman, V. Raman, R. Sidle, M. Spilchen, A. Storm, Y. Tian, P. Tozun, D. Zilio, M. Huras, C. Mohan, F. Ozcan, H. Pirahesh IBM Research, IBM Analytics

2 Apps emit tons of events IOT Mobile commerce Today s DBMS is too slow for this firehose And too costly (storage) Events happen in the real-world not tied to transaction commit Place Order Withdraw Cash Events always have asterisks Concurrent events Event ordering done later wearables

3 Apps emit tons of events IOT Mobile commerce wearables Today s DBMS is too slow for this firehose Multi-master, And too costly (storage) disconnected operation Events happen in the real-world not tied to transaction commit Place Order Withdraw Cash Events always have asterisks Concurrent events Event ordering done later Events need Availability and Consistency Today: Consistency achieved via application logic Compensation, apologies, coupons, Weak atomicity and durability Growing pressure for DBMS to give both Availability and Consistency Due to globalization e.g., credit cards

4 Apps emit tons of events IOT Mobile commerce wearables Events need Avail. and Consistency Events need HTAP Today s DBMS is too slow for this firehose Multi-master, Mobile commerce And too costly (storage) disconnected operation Retail: inventory analysis, shipping time Events happen in the real-world Today: Consistency analysisachieved via not tied to transaction commit application logic Securities trading: global risk analysis Withdraw Place Order Compensation, apologies, Cash Margin rules coupons, Events always have asterisks Weak atomicity Complex and analysis durability within transaction Concurrent events Touching lots of rows Growing pressure for DBMS to give Event ordering done later both Availability Handled Consistency poorly by both 2PL and OCC Due to globalization Analysis involves far more than SQL e.g., credit cards Graph, machine learning,

5 Apps emit tons of events IOT Mobile commerce WildFire Goals wearables Events need Avail. and Consistency Events need HTAP Today s DBMS is too slow for this firehose Multi-master, Mobile commerce And too costly (storage) disconnected operation Peak transaction speed Retail: Multi-Master inventory analysis, and ACID shipping time Events Inserts/updates: happen in the keep real-world up with bandwidth Today: Consistency of analysis Commit achieved is local via (no consensus) durability not tied to mechanism transaction (today: commit 1e6/s/node) application logic Securities High-value trading: events global can risk wait analysis for Full indexing: Withdraw Place Order keep up w. random access Compensation, speed apologies, Cash coupons, Complex conflict analysis resolution within transaction (after commit ) (hash tables + atomics on SSD à NVRAM) Events Fully versioned always have asterisks Weak atomicity Handled durability poorly by both 2PL and OCC Concurrent events HTAP Growing pressure for DBMS to give Open Event Format ordering done later In-transaction analytics both Availability and Consistency Animation All data groomed WF targets to Parquet 1M format Analytics over snapshot Due to globalization xsacs/s/node on shared storage (run at bandwidth of (1s or 10mins) e.g., credit cards durable Directly medium accessible / network) by analytics platforms Higher throughput and (e.g., Spark) more economical scaleout

6 Wildfire architecture Applications analytics can tolerate slightly stale data requires most recent data high-volume transactions wildfire engine wildfire engine wildfire engine SSD/NVM wildfire engine SSD/NVM shared file system

7 Data lifecycle OLTP nodes postgroom groom ORGANIZED zone TIME (PBs of data) HTAP (see latest: snapshot isolation) 1-sec old snapshot Optimized snapshot (10 mins stale) Analytics nodes GROOMED zone (~10 mins) LIVE zone (~1sec) ML, etc (Spark) BI Snapshot Lookups

8 xsacs Live Zone: Thin Commit per xsac logs (uncommitted) log (committed) replicate xsacs What happens at Commit 1. append xsac deltas (Ins/Del/Upd) to common log; replicated in background -- everything is an upsert: key, (values)* -- no synchronous conflict resolution 2. flush to local SSD 3. high-value xsacs wait for grooming (to timestamp the xsac and resolve conflicts) -- can time-out Driven by speed No tracking down prior versions No indexing No waiting for consensus with other nodes multiple versions for same key can coexist -- queries pick right version based on their xsac snapshot

9 xsacs Grooming (Live à Groomed zone) per xsac logs (uncommitted) replicate xsacs log (committed) groom Runs distributed consensus to timestamp the xsacs (pick serialization order) take quorum-visible deltas, form data blocks, and publish to shared file system Add begints field to each row: (groomts localtime nodeid) Conflicts and constraints resolved lazily (including logical rollback) No assumption about Clock synchronization Partitioning / failures (multiple groomers possible) Details offline

10 Postgrooming Organize data so that Queries can run fast (and deal with immutable storage!) Resolve conflicts (and stamp xsacs with resolution/rollback status) Compute endtime and prevrid Partition data (along multiple dimensions) Maintain primary indexes, secondary indexes, and synopses Continual refinement done in background - big challenge is supporting concurrent groom and queries CURRENT Partitions Groomed Block Log HISTORY Partitions ORGANIZED Zone GROOMED Zone LIVE Zone

11 Postgrooming: index maintenance Primary: maps key hash à RID [+ include columns] Secondary: maps key hash à pkey [+ TSN hint] Works in background to add groomed records to indexes Index is variant of LSM tree Merging in background Lives in multiple tiers: memory, SSD, shared storage, and purged Multiple versions of each key live in index 11

12 Postgrooming: computing endts At groom, each row has a begints - but no endts Without postgrooming, every table scan must group-by on primary key to pick appropriate version Postgroom picks groomed rows and assigns endts 1. Massive set intersection: PrimaryKeyIndex (keyàlatestrid) with RecentlyGroomed prior versions can be arbitrarily far back! 2. Squeeze in endts into data blocks (with concurrent readers!)

13 Postgrooming: resolving transaction status (single-shard) ReadSet tracking is pessimistic, especially with complex queries Eg: currentinventory ß select sum(..) from ledger where productid=_ if (currentinventory > 2) insert into ledger values (-1, productid, ); # buy one item Constraint-based resolution Transaction Type1: { read* ; fullyspecifiedwrite*; If (trigger) { rollback or other action }} ATM withdrawal, Securities trading (higher-granularity checks) Transaction Type2: { read* ; if (trigger) { fullyspecifiedwrite* } } Submit order Trigger Condition is checked as a continuous query (incrementally feed new deltas) ReadSet-based resolution {Read*; write*; if (trigger) {rollback or other action}} (trigger readset changed since query snapshot) is checked as a continuous query More general transactions (eg RWRW) hard to check incrementally fall back to traditional readset tracking

14 Postgrooming: resolving transaction status (multi-shard) Each transaction can produce multiple deltas, spread across groom cycles No 2PC Each transaction stamped with its delta count Resolve only considers a delta if all deltas of that transaction are available

15 Concluding Remarks OLTP/OLAP separation is going away DBMS needs to be much faster, and stop controlling the data format, and stop controlling the data storage, and stop controlling the kinds of analytics DBMS can be the manager for event data à V V LDB Thank you

16 BACKUP

10. Replication. CSEP 545 Transaction Processing Philip A. Bernstein. Copyright 2003 Philip A. Bernstein. Outline

10. Replication. CSEP 545 Transaction Processing Philip A. Bernstein. Copyright 2003 Philip A. Bernstein. Outline 10. Replication CSEP 545 Transaction Processing Philip A. Bernstein Copyright 2003 Philip A. Bernstein 1 Outline 1. Introduction 2. Primary-Copy Replication 3. Multi-Master Replication 4. Other Approaches