Pengaruh Mesin Pencarian dan Arsitektur Dalam Mengoptimumkan Kinerja dan Skalabilitas Sistem Katalog Digital Perpustakaan

Abstract

Sistem katalog digital perpustakaan dituntut untuk mengelola data bibliografis berskala besar sekaligus memberikan layanan pencarian yang cepat dan relevan. Penelitian ini mengevaluasi pengaruh konfigurasi mesin pencarian berbasis Apache Solr dan pilihan arsitektur sistem terhadap kinerja dan skalabilitas sistem katalog digital, serta mengembangkan strategi pencarian adaptif berdasarkan karakteristik kueri pengguna, menggunakan dataset 1.527.259 rekaman bibliografis dari sistem INLIS Perpustakaan Nasional Republik Indonesia (Perpusnas). Lima pendekatan pencarian Apache Solr dievaluasi, terdiri dari empat konfigurasi tokenizer (text_standart, text_edgengram, text_ngram, text_typo_tolerant) dan satu teknik pencarian fuzzy berbasis operator kueri, menggunakan 30 kueri original dan 30 kueri typo yang diukur dengan metrik precision, recall, dan F1-score, kemudian dianalisis dengan uji One-Way ANOVA. Perbandingan arsitektur monolitik dan Microservices yang dibangun dengan Laravel 11 dilakukan menggunakan Apache JMeter pada tiga tingkat beban (200, 500, dan 1.000 concurrent users), mencakup skenario CRUD, pencarian katalog, dan beban campuran. Uji skalabilitas dilakukan melalui Vertical Scaling dan horizontal scaling. Hasil evaluasi menunjukkan bahwa perbedaan konfigurasi tokenizer berpengaruh signifikan terhadap performa pencarian, terutama pada kueri typo (p = 1,14 × 10?5) dan kueri campuran (p = 5,09 × 10?5). Strategi pencarian adaptif yang diusulkan, yang mengombinasikan konfigurasi text_standart, text_ngram, dan fuzzy search secara dinamis berdasarkan panjang kueri, menghasilkan precision 0,735, recall 0,993, dan F1-score 0,795 — meningkat 21,6% dibandingkan konfigurasi terbaik tunggal (text_ngram, F1 = 0,654). Temuan penting lainnya adalah kueri dua kata merupakan titik paling kritis bagi sistem, bukan kueri terpanjang, karena ambiguitas pada level tersebut menjadi penyebab utama penurunan performa. Dari sisi arsitektur, arsitektur monolitik menunjukkan keunggulan waktu respons pada beban rendah (200 pengguna, median 2.707 ms) dibandingkan microservices (3.220 ms), karena tidak adanya overhead komunikasi antar layanan. Sebaliknya, arsitektur Microservices secara konsisten menunjukkan throughput lebih tinggi dan waktu respons lebih rendah pada operasi pencarian katalog dan beban campuran di seluruh tingkat beban, dengan selisih mencapai 4.998 ms pada 1.000 pengguna. Pada uji Vertical Scaling, penambahan sumber daya menghasilkan speedup 1,29× pada Microservices dan 1,25× pada monolitik. Uji horizontal scaling pada Microservices menghasilkan speedup hingga 1,58× pada empat instance, meskipun pertumbuhan tidak linear akibat bottleneck pada komponen shared data layer dan gateway. Analisis cost-benefit menunjukkan bahwa meskipun arsitektur microservices membutuhkan infrastruktur lebih besar, efisiensi throughput per vCPU-nya tetap lebih tinggi (15,9 req/s/vCPU dibandingkan 14,8 req/s/vCPU pada monolitik). Simulasi proyeksi biaya sepuluh tahun menunjukkan microservices berpotensi lebih efisien secara kumulatif dalam jangka panjang karena mampu melakukan horizontal scaling tanpa lompatan biaya migrasi instance seperti pada monolitik, serta memberikan manfaat tangible dan intangible yang lebih besar, meski monolitik tetap unggul pada kesederhanaan pengelolaan operasional. Kombinasi strategi pencarian adaptif dengan migrasi bertahap menuju arsitektur microservices, khususnya pada layanan pencarian katalog, direkomendasikan sebagai pendekatan yang paling sesuai untuk sistem katalog digital perpustakaan berskala besar seperti Perpusnas.

Digital library catalog systems are required to manage large-scale bibliographic data while delivering fast and relevant search services. This study evaluated the effect of Apache Solr search engine configuration and system architecture on the performance and scalability of a digital library catalog system, and developed an Adaptive search strategy based on query characteristics, using a dataset of 1,527,259 bibliographic records from the INLIS system of the National Library of Indonesia (Perpusnas). Five Apache Solr search approaches were evaluated, consisting of four tokenizer configurations (text_standart, text_edgengram, text_ngram, text_typo_tolerant) and one fuzzy search technique based on a query-time operator, measured using 30 original and 30 typo queries evaluated with precision, recall, and F1-score metrics, then analyzed through One-Way ANOVA tests. A comparison of monolithic and microservices architectures, both built on Laravel 11, was conducted using Apache JMeter under loads of 200, 500, and 1,000 concurrent users, covering CRUD operations, catalog search, and mixed-load scenarios. Scalability was assessed through vertical and horizontal scaling experiments. Results showed that differences in tokenizer configuration significantly affected search performance, particularly for typo queries (p = 1.14 × 10?5) and mixed queries (p = 5.09 × 10?5). The proposed Adaptive search strategy, which dynamically combines text_standart, text_ngram, and fuzzy search based on query length, achieved a precision of 0.735, recall of 0.993, and F1-score of 0.795 — a 21.6% improvement over the best single configuration (text_ngram, F1 = 0.654). A key finding was that two-word queries posed the greatest challenge to the system, rather than the longest queries, due to increased ambiguity at that complexity level. Regarding architecture, the monolithic approach delivered lower response times under low load (200 users: median 2,707 ms) compared to microservices (3,220 ms), due to the absence of inter-service communication overhead. Conversely, the microservices architecture consistently showed higher throughput and lower response times for catalog search and mixed-load operations across all load levels, with a difference of up to 4,998 ms at 1,000 concurrent users. In the Vertical Scaling test, additional resources yielded a speedup of 1.29× for microservices and 1.25× for monolithic. Horizontal scaling of microservices achieved speedups of up to 1.58× with four instances, although growth was sub-linear due to bottlenecks at the shared data layer and gateway components. Cost-benefit analysis showed that the microservices architecture required higher development, start-up, and annual maintenance costs than the monolithic architecture due to the complexity of managing independent services, but was more cost-efficient in terms of annual infrastructure costs and throughput per vCPU (15.9 req/s/vCPU compared to 14.8 req/s/vCPU for monolithic). A ten-year cumulative cost projection simulation indicated that microservices could be more cost-efficient in the long term, as it allows horizontal scaling without the instance-migration cost jumps required by monolithic architecture, while also delivering greater tangible and intangible benefits, although the monolithic architecture remained superior in terms of operational simplicity. The combination of the Adaptive search strategy with a phased migration toward microservices architecture, particularly for the catalog search service, is recommended as the most suitable approach for large-scale digital library catalog systems such as Perpusnas.