Sparser: Faster Parsing of Unstructured Data Formats in Apache Spark with Firas Abuzaid and Shoumik Palkar
2 likes2,149 views
The document introduces 'sparser', a method for accelerating the querying of large unstructured datasets by filtering data before parsing, significantly improving efficiency. It outlines how sparser integrates with existing data sources in Spark, utilizing raw filters to reduce the computational bottleneck associated with parsing. Performance results demonstrate up to 22x speedup over existing parsers and emphasize sparser's potential for exploratory analytics.
![2. Measure RFs and Parser
Passthrough rates of RFs are non-independent! But too
expensive to rate of each cascade (combinatorial search
space)
Solution: Store rates as bitmaps.
Pr[a] ∝ 1s in Bitmap of RF a: 0011010101
Pr[b] ∝ 1s in Bitmap of RF b: 0001101101
Pr[a,b] ∝ 1s in Bitwise-& of a, b: 0001000101
foreach sampled record i:
foreach RF j:
BitMap[i][j] = 1 if RF j passes on i](https://image.slidesharecdn.com/4firasabuzaidshoumikpalkar-180612014118/85/Sparser-Faster-Parsing-of-Unstructured-Data-Formats-in-Apache-Spark-with-Firas-Abuzaid-and-Shoumik-Palkar-32-320.jpg)
![Probability and cost of
executing full parser
Probability and cost
of executing RF i
Cost of Cascade
with RFs 0,1,…,R
3. Choosing Best Cascade
𝐶" = $ Pr [ 𝑒𝑥𝑒𝑐𝑢𝑡𝑒.]
.∈"
×𝑐. + Pr 𝑒𝑥𝑒𝑐𝑢𝑡𝑒45678 ×𝑐45678
P
Trum
utinObam
PD Obam
PD
Pr 𝑒𝑥𝑒𝑐𝑢𝑡𝑒96:; = 1
Pr 𝑒𝑥𝑒𝑐𝑢𝑡𝑒:=.> = Pr [Trum]
Pr 𝑒𝑥𝑒𝑐𝑢𝑡𝑒BC5; = Pr [¬Trum] + Pr[Trum,¬utin]
Pr 𝑒𝑥𝑒𝑐𝑢𝑡𝑒45678 = Pr [¬Trum,Obam] + Pr[Trum,utin] + Pr[Trum,¬utin,Obam]
Choose cascade with minimum cost: Sparser considers up to min(32,
# ORs) RFs and considers up to depth of 4 cascades.](https://image.slidesharecdn.com/4firasabuzaidshoumikpalkar-180612014118/85/Sparser-Faster-Parsing-of-Unstructured-Data-Formats-in-Apache-Spark-with-Firas-Abuzaid-and-Shoumik-Palkar-34-320.jpg)
Sparser: Faster Parsing of Unstructured Data Formats in Apache Spark with Firas Abuzaid and Shoumik Palkar
- 1. Sparser: Fast Analytics on Raw Data by Avoiding Parsing Shoumik Palkar, Firas Abuzaid, Peter Bailis, Matei Zaharia
- 2. Motivation Bigger unstructured datasets and faster hardware. Time Speed Time DataVolume
- 3. Parsing unstructured data before querying it is often very slow.
- 4. Today: Spark Data Sources API Spark Core Engine Push part of query into the data source.
- 5. Today: Spark Data Sources API + E.g., column pruning directly in Parquet data loader - Little support for unstructured formats (e.g., can’t avoid JSON parsing) Spark Core Engine Data Source API Push part of query into the data source.
- 6. Parsing: A Computational Bottleneck Example: Existing state-of-the- art JSON parsers 100x slower than scanning a string! 113836 56796 360 0 50000 100000 150000 RapidJSON Parsing Mison Index Contruction String Scan Cycles/5KBRecord State-of-the-art Parsers *Similar results on binary formats like Avro and Parquet Parsing seems to be a necessary evil: how do we get around doing it?
- 7. Key Opportunity: High Selectivity High selectivity especially true for exploratory analytics. 0 0.2 0.4 0.6 0.8 1 1.E-09 1.E-05 1.E-01 CDF Selectivity Databricks Censys 40% of customer Spark queries at Databricks select < 20% of data 99% of queries in Censys select < 0.001% of data Data Source API provides access to query filters at data source!
- 8. How can we exploit high selectivity to accelerate parsing?
- 9. Sparser: Filter Before You Parse Raw DataFilter Raw DataFilter Raw DataFilter Raw DataParse Raw DataParse Today: parse full input à slow! Sparser: Filter before parsing first using fast filtering functions with false positives, but no false negatives
- 10. Demo Count Tweets where text contains “Trump” and “Putin”
- 11. Sparser in Spark SQL Spark Core Engine Data Source API val df = spark.read.format(“json”)
- 12. Sparser in Spark SQL Spark Core Engine Data Source API Sparser Filtering Engine val df = spark.read.format(“edu.stanford.sparser.json”) Sparser Data Source Reader (Also supports Avro, Parquet!)
- 13. Sparser Overview
- 14. Sparser Overview Raw Filter (“RF”): filtering function on a bytestream with false positives but no false negatives. Use an optimizer to combine RFs into a cascade. But is it in the “text” field? Is the full word “Trump”? 0x54 0x72 0x75 0x6d 0x70 T r u m b RF “TRUM” Pass! Other RFs Example: Tweets WHERE text contains “Trump” Full Parser
- 15. Key Challenges in Using RFs 1. How do we implement the RFs efficiently? 2. How do we efficiently choose the RFs to maximize parsing throughput for a given query and dataset? Rest of this Talk • Sparser’s API • Sparser’s Raw Filter Designs (Challenge 1) • Sparser’s Optimizer: Choosing a Cascade (Challenge 2) • Performance Results on various file formats (e.g., JSON, Avro)
- 16. Sparser API
- 17. Sparser API Filter Example Exact String Match WHERE user.name = “Trump” WHERE likes = 5000
- 18. Sparser API Filter Example Exact String Match WHERE user.name = “Trump” WHERE likes = 5000 Contains String WHERE text contains “Trum”
- 19. Sparser API Filter Example Exact String Match WHERE user.name = “Trump” WHERE likes = 5000 Contains String WHERE text contains “Trum” Contains Key WHERE user.url != NULL
- 20. Sparser API Filter Example Exact String Match WHERE user.name = “Trump” WHERE likes = 5000 Contains String WHERE text contains “Trum” Contains Key WHERE user.url != NULL Conjunctions WHERE user.name = “Trump” AND user.verified = true Disjunctions WHERE user.name = “Trump” OR user.name = “Obama”
- 21. Sparser API Filter Example Exact String Match WHERE user.name = “Trump” WHERE likes = 5000 Contains String WHERE text contains “Trum” Contains Key WHERE user.url != NULL Conjunctions WHERE user.name = “Trump” AND user.verified = true Disjunctions WHERE user.name = “Trump” OR user.name = “Obama” Currently does not support numerical range-based predicates.
- 22. Challenge 1: Efficient RFs
- 23. Raw Filters in Sparser SIMD-based filtering functions that pass or discard a record by inspecting raw bytestream 113836 56796 360 0 50000 100000 150000 RapidJSON Parsing Mison Index Contruction Raw Filter Cycles/5KBRecord
- 24. Example RF: Substring Search Search for a small (e.g., 2, 4, or 8-byte) substring of a query predicate in parallel using SIMD On modern CPUs: compare 32 characters in parallel in ~4 cycles (2ns). Length 4 Substring packed in SIMD register Input : “ t e x t ” : “ I j u s t m e t M r . T r u m b ! ! ! ” Shift 1 : T r u m T r u m T r u m T r u m T r u m T r u m T r u m - - - - - - - Shift 2 : - T r u m T r u m T r u m T r u m T r u m T r u m T r u m - - - - - - Shift 3 : - - T r u m T r u m T r u m T r u m T r u m T r u m T r u m - - - - - Shift 4 : - - - T r u m T r u m T r u m T r u m T r u m T r u m T r u m - - - - Example query: text contains “Trump” False positives (found “Trumb” by accident), but no false negatives (No “Trum” ⇒ No “Trump”) Other RFs also possible! Sparser selects them agnostic of implementation.
- 25. Key-Value Search RF Searches for key, and if key is found, searches for value until some stopping point. Searches occur with SIMD. Only applicable for exact matches Useful for queries with common substrings (e.g., favorited=true) Key: name Value: Trump Delimiter: , Ex 1: “name”: “Trump”, Ex 2: “name”: “Actually, Trump” Ex 3: “name”: “My name is Trump” (Pass) (Fail) (Pass, False positive) Second Example would result in false negative if we allow substring matches Other RFs also possible! Sparser selects them agnostic of implementation.
- 26. Challenge 2: Choosing RFs
- 27. Choosing RFs To decrease false positive rate, combine RFs into a cascade Sparser uses an optimizer to choose a cascade Raw DataFilter Raw DataFilter Raw DataFilter Raw DataParse Raw DataFilter Raw DataParse Raw DataFilter Raw DataParse Raw DataFilter Raw DataFilter vs. vs.
- 28. Sparser’s Optimizer Step 2: Measure Params on Sample of Records raw bytestream 0010101000110101 RF 1 RF fails RF 2 X X RF fails Filtered bytes sent to full parser Step 4: Apply Chosen Cascade for sampled records: 1 = passed 0 = failed Step 3: Score and Choose Cascade … (name = "Trump" AND text contains "Putin") Step 1: Compile Possible RFs from predicates RF1: "Trump" RF2: "Trum" RF3: "rump" RF4: "Tr" … RF5: "Putin" C (RF1) = 4 C (RF1àRF2) = 1 C (RF2àRF3) = 6 C (RF1àRF3) = 9 … S1 S2 S3 RF 1 0 1 1 RF 2 1 0 1
- 29. Sparser’s Optimizer: Configuring RF Cascades Three high level steps: 1. Convert a query predicate to a set of RFs 2. Measure the passthrough rate and runtime of each RF and the runtime of the parser on sample of records 3. Minimize optimization function to find min-cost cascade
- 30. 1. Converting Predicates to RFs Running Example (name = “Trump” AND text contains “Putin”) OR name = “Obama” Possible RFs* Substring Search: Trum, Tr, ru, …, Puti, utin, Pu, ut, …, Obam, Ob, … * Translation from predicate to RF is format-dependent: examples are for JSON data
- 31. 2. Measure RFs and Parser Evaluate RFs and parser on a sample of records. *Passthrough rates of RFs are non-independent! But too expensive to measure rate of each cascade (combinatorial search space) RF RF Parser Passthrough rates* and runtimes
- 32. 2. Measure RFs and Parser Passthrough rates of RFs are non-independent! But too expensive to rate of each cascade (combinatorial search space) Solution: Store rates as bitmaps. Pr[a] ∝ 1s in Bitmap of RF a: 0011010101 Pr[b] ∝ 1s in Bitmap of RF b: 0001101101 Pr[a,b] ∝ 1s in Bitwise-& of a, b: 0001000101 foreach sampled record i: foreach RF j: BitMap[i][j] = 1 if RF j passes on i
- 33. 3. Choosing Best Cascade Running Example (name = “Trump” AND text contains “Putin”) OR name = “Obama” P Trum utinObam PD Obam PD 1. Valid cascade 2. Valid cascade utin Obam PD P 3. Invalid cascade (utin must consider Obam when it fails) Trum utinObam PD PD Pass (right branch)Fail (left branch)Parse PDiscard D Cascade cost computed by traversing tree from root to leaf. Each RF has a runtime/passthrough rate we measured in previous step.
- 34. Probability and cost of executing full parser Probability and cost of executing RF i Cost of Cascade with RFs 0,1,…,R 3. Choosing Best Cascade 𝐶" = $ Pr [ 𝑒𝑥𝑒𝑐𝑢𝑡𝑒.] .∈" ×𝑐. + Pr 𝑒𝑥𝑒𝑐𝑢𝑡𝑒45678 ×𝑐45678 P Trum utinObam PD Obam PD Pr 𝑒𝑥𝑒𝑐𝑢𝑡𝑒96:; = 1 Pr 𝑒𝑥𝑒𝑐𝑢𝑡𝑒:=.> = Pr [Trum] Pr 𝑒𝑥𝑒𝑐𝑢𝑡𝑒BC5; = Pr [¬Trum] + Pr[Trum,¬utin] Pr 𝑒𝑥𝑒𝑐𝑢𝑡𝑒45678 = Pr [¬Trum,Obam] + Pr[Trum,utin] + Pr[Trum,¬utin,Obam] Choose cascade with minimum cost: Sparser considers up to min(32, # ORs) RFs and considers up to depth of 4 cascades.
- 36. Sparser in Spark SQL Spark Core Engine Data Source API Sparser Filtering Engine
- 37. Sparser in Spark SQL Spark Core Engine Data Source API Sparser Filtering Engine SQL Query to Candidate Raw Filters Native HDFS File Reader Sparser Optimizer + Filtering Raw Data DataFrame
- 39. Evaluation Setup Datasets Censys - Internet port scan data used in network security Tweets - collected using Twitter Stream API Distributed experiments on 20-node cluster with 4 Broadwell vCPUs/26GB of memory, locally attached SSDs (Google Compute Engine) Single node experiments on Intel Xeon E5-2690 v4 with 512GB memory
- 40. Queries Twitter queries from other academic work. Censys queries sampled randomly from top-3000 queries. PCAP and Bro queries from online tutorials/blogs.
- 41. Results: Accelerating End-to-End Spark Jobs 0 200 400 600 Disk Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8 Q9 Runtime(seconds) Spark Spark + Sparser Query Only Censys queries on 652GB of JSON data: up to 4x speedup by using Sparser. 1.1 1.3 0.7 1.4 0 20 40 60 80 Disk Q1 Q2 Q3 Q4 Runtime (seconds) Spark Spark + Sparser Query Only Twitter queries on 68GB of JSON data: up to 9x speedup by using Sparser.
- 42. Results: Accelerating Existing JSON Parsers 1 10 100 1 2 3 4 5 6 7 8 9 Runtime(s,log10) Sparser + RapidJSON Sparser + Mison RapidJSON Mison Censys queries compared against two state-of-the-art parsers (Mison based on SIMD): Sparser accelerates them by up to 22x.
- 43. Results: Accelerating Binary Format Parsing 0 0.5 1 1.5 Q1 Q2 Q3 Q4 Runtime (seconds) avro-c Sparser + avro-c 0 0.2 Q1 Q2 Q3 Q4 Runtime (seconds) parquet-cpp Sparser + parquet-cpp Sparser accelerates Avro (above) and Parquet (below) based queries by up to 5x and 4.3x respectively.
- 44. Results: Domain-Specific Tasks 0 1 2 3 4 0.01 0.1 1 10 100 1000 P1 P2 P3 P4 Speedupover libpcap Runtime(s,log10) Sparser libpcap Tshark tcpdump Sparser accelerates packet parsing by up to 3.5x compared to standard tools.
- 45. Results: Sparser’s Sensitivity to Selectivity 0 20 40 60 80 100 120 0 0.2 0.4 0.6 0.8 1 Runtime(seconds) Selectivity RapidJSON Sparser + RapidJSON Mison Sparser + Mison Sparser’s performance degrades gracefully as the selectivity of the query decreases.
- 46. Results: Nuts and Bolts 1 10 100 0 200 400 Runtime(s,log10) Cascade Sparser picks a cascade within 5% of the optimal one using its sampling-based measurement and optimizer. Picked by Sparser 100 1000 10000 0 100 200 ParseThroughput (MB/s) File Offset (100s of MB) With Resampling No Resampling Resampling to calibrate the cascade can improve end-to-end parsing time by up to 20x.
- 47. Conclusion Sparser: • Uses raw filters to filter before parsing • Selects a cascade of raw filters using an efficient optimizer • Delivers up to 22x speedups over existing parsers • Will be open source soon! [email protected] [email protected] Questions or Comments? Contact Us!